Design and Analysis of Experiments

Design and Analysis of Experiments Design and Analysis of Experiments Volume 2 Advanced Experimental Design KLAUS HI...

Author: Klaus Hinkelmann | Oscar Kempthorne

451 downloads 3963 Views 4MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form

DOWNLOAD PDF

Design and Analysis of Experiments

Design and Analysis of Experiments Volume 2 Advanced Experimental Design

KLAUS HINKELMANN Virginia Polytechnic Institute and State University Department of Statistics Blacksburg, VA

OSCAR KEMPTHORNE Iowa State University Department of Statistics Ames, IA

A JOHN WILEY & SONS, INC., PUBLICATION

Copyright  2005 by John Wiley & Sons, Inc. All rights reserved. Published by John Wiley & Sons, Inc., Hoboken, New Jersey. Published simultaneously in Canada. No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, scanning, or otherwise, except as permitted under Section 107 or 108 of the 1976 United States Copyright Act, without either the prior written permission of the Publisher, or authorization through payment of the appropriate per-copy fee to the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, 978-750-8400, fax 978-646-8600, or on the web at www.copyright.com. Requests to the Publisher for permission should be addressed to the Permissions Department, John Wiley & Sons, Inc., 111 River Street, Hoboken, NJ 07030, (201) 748-6011, fax (201) 748-6008. Limit of Liability/Disclaimer of Warranty: While the publisher and author have used their best efforts in preparing this book, they make no representations or warranties with respect to the accuracy or completeness of the contents of this book and specifically disclaim any implied warranties of merchantability or fitness for a particular purpose. No warranty may be created or extended by sales representatives or written sales materials. The advice and strategies contained herein may not be suitable for your situation. You should consult with a professional where appropriate. Neither the publisher nor author shall be liable for any loss of profit or any other commercial damages, including but not limited to special, incidental, consequential, or other damages. For general information on our other products and services please contact our Customer Care Department within the U.S. at 877-762-2974, outside the U.S. at 317-572-3993 or fax 317-572-4002. Wiley also publishes its books in a variety of electronic formats. Some content that appears in print, however, may not be available in electronic format. Library of Congress Cataloging-in-Publication Data is available. ISBN 0-471-55177-5 Printed in the United States of America. 10 9 8 7 6 5 4 3 2 1

Contents

Preface 1

xix

General Incomplete Block Design 1.1 1.2 1.3

1.4 1.5 1.6

1.7

1.8

1

Introduction and Examples, 1 General Remarks on the Analysis of Incomplete Block Designs, 3 The Intrablock Analysis, 4 1.3.1 Notation and Model, 4 1.3.2 Normal and Reduced Normal Equations, 5 1.3.3 The C Matrix and Estimable Functions, 7 1.3.4 Solving the Reduced Normal Equations, 8 1.3.5 Estimable Functions of Treatment Effects, 10 1.3.6 Analyses of Variance, 12 Incomplete Designs with Variable Block Size, 13 Disconnected Incomplete Block Designs, 14 Randomization Analysis, 16 1.6.1 Derived Linear Model, 16 1.6.2 Randomization Analysis of ANOVA Tables, 18 Interblock Information in an Incomplete Block Design, 23 1.7.1 Introduction and Rationale, 23 1.7.2 Interblock Normal Equations, 23 1.7.3 Nonavailability of Interblock Information, 27 Combined Intra- and Interblock Analysis, 27 1.8.1 Combining Intra- and Interblock Information, 27 1.8.2 Linear Model, 27 1.8.3 Normal Equations, 28 1.8.4 Some Special Cases, 31 v

vi

CONTENTS

1.9

1.10

1.11

1.12

1.13

1.14

2

Relationships Among Intrablock, Interblock, and Combined Estimation, 31 1.9.1 General Case, 32 1.9.2 Case of Proper, Equireplicate Designs, 34 Estimation of Weights for the Combined Analysis, 36 1.10.1 Yates Procedure, 37 1.10.2 Properties of Combined Estimators, 38 Maximum-Likelihood Type Estimation, 39 1.11.1 Maximum-Likelihood Estimation, 39 1.11.2 Restricted Maximum-Likelihood Estimation, 40 Efficiency Factor of an Incomplete Block Design, 43 1.12.1 Average Variance for Treatment Comparisons for an IBD, 43 1.12.2 Definition of Efficiency Factor, 45 1.12.3 Upper Bound for the Efficiency Factor, 47 Optimal Designs, 48 1.13.1 Information Function, 48 1.13.2 Optimality Criteria, 49 1.13.3 Optimal Symmetric Designs, 50 1.13.4 Optimality and Research, 50 Computational Procedures, 52 1.14.1 Intrablock Analysis Using SAS PROC GLM, 52 1.14.2 Intrablock Analysis Using the Absorb Option in SAS PROC GLM, 58 1.14.3 Combined Intra- and Interblock Analysis Using the Yates Procedure, 61 1.14.4 Combined Intra- and Interblock Analysis Using SAS PROC MIXED, 63 1.14.5 Comparison of Estimation Procedures, 63 1.14.6 Testing of Hypotheses, 66

Balanced Incomplete Block Designs 2.1 2.2 2.3 2.4

2.5

Introduction, 71 Definition of the BIB Design, 71 Properties of BIB Designs, 72 Analysis of BIB Designs, 74 2.4.1 Intrablock Analysis, 74 2.4.2 Combined Analysis, 76 Estimation of ρ, 77

71

vii

CONTENTS

2.6 2.7

2.8

3

Construction of Balanced Incomplete Block Designs 3.1 3.2

3.3

3.4 4

Significance Tests, 79 Some Special Arrangements, 89 2.7.1 Replication Groups Across Blocks, 89 2.7.2 Grouped Blocks, 91 2.7.3 α-Resolvable BIB Designs with Replication Groups Across Blocks, 96 Resistant and Susceptible BIB Designs, 98 2.8.1 Variance-Balanced Designs, 98 2.8.2 Definition of Resistant Designs, 99 2.8.3 Characterization of Resistant Designs, 100 2.8.4 Robustness and Connectedness, 103

Introduction, 104 Difference Methods, 104 3.2.1 Cyclic Development of Difference Sets, 104 3.2.2 Method of Symmetrically Repeated Differences, 107 3.2.3 Formulation in Terms of Galois Field Theory, 112 Other Methods, 113 3.3.1 Irreducible BIB Designs, 113 3.3.2 Complement of BIB Designs, 113 3.3.3 Residual BIB Designs, 114 3.3.4 Orthogonal Series, 114 Listing of Existing BIB Designs, 115

Partially Balanced Incomplete Block Designs 4.1 4.2

4.3

4.4

104

Introduction, 119 Preliminaries, 119 4.2.1 Association Scheme, 120 4.2.2 Association Matrices, 120 4.2.3 Solving the RNE, 121 4.2.4 Parameters of the Second Kind, 122 Definition and Properties of PBIB Designs, 123 4.3.1 Definition of PBIB Designs, 123 4.3.2 Relationships Among Parameters of a PBIB Design, 125 Association Schemes and Linear Associative Algebras, 127 4.4.1 Linear Associative Algebra of Association Matrices, 127

119

viii

CONTENTS

4.5

4.6

4.7

5

4.4.2 Linear Associative Algebra of P Matrices, 128 4.4.3 Applications of the Algebras, 129 Analysis of PBIB Designs, 131 4.5.1 Intrablock Analysis, 131 4.5.2 Combined Analysis, 134 Classification of PBIB Designs, 137 4.6.1 Group-Divisible (GD) PBIB(2) Designs, 137 4.6.2 Triangular PBIB(2) Designs, 139 4.6.3 Latin Square Type PBIB(2) Designs, 140 4.6.4 Cyclic PBIB(2) Designs, 141 4.6.5 Rectangular PBIB(3) Designs, 142 4.6.6 Generalized Group-Divisible (GGD) PBIB(3) Designs, 143 4.6.7 Generalized Triangular PBIB(3) Designs, 144 4.6.8 Cubic PBIB(3) Designs, 146 4.6.9 Extended Group-Divisible (EGD) PBIB Designs, 147 4.6.10 Hypercubic PBIB Designs, 149 4.6.11 Right-Angular PBIB(4) Designs, 151 4.6.12 Cyclic PBIB Designs, 153 4.6.13 Some Remarks, 154 Estimation of ρ for PBIB(2) Designs, 155 4.7.1 Shah Estimator, 155 4.7.2 Application to PBIB(2) Designs, 156

Construction of Partially Balanced Incomplete Block Designs 5.1

5.2

5.3

5.4

Group-Divisible PBIB(2) Designs, 158 5.1.1 Duals of BIB Designs, 158 5.1.2 Method of Differences, 160 5.1.3 Finite Geometries, 162 5.1.4 Orthogonal Arrays, 164 Construction of Other PBIB(2) Designs, 165 5.2.1 Triangular PBIB(2) Designs, 165 5.2.2 Latin Square PBIB(2) Designs, 166 Cyclic PBIB Designs, 167 5.3.1 Construction of Cyclic Designs, 167 5.3.2 Analysis of Cyclic Designs, 169 Kronecker Product Designs, 172 5.4.1 Definition of Kronecker Product Designs, 172

158

ix

CONTENTS

5.5

5.6 6

5.4.2 Properties of Kronecker Product Designs, 172 5.4.3 Usefulness of Kronecker Product Designs, 177 Extended Group-Divisible PBIB Designs, 178 5.5.1 EGD-PBIB Designs as Kronecker Product Designs, 178 5.5.2 Method of Balanced Arrays, 178 5.5.3 Direct Method, 180 5.5.4 Generalization of the Direct Method, 185 Hypercubic PBIB Designs, 187

More Block Designs and Blocking Structures 6.1 6.2

6.3 6.4

6.5

6.6

Introduction, 189 Alpha Designs, 190 6.2.1 Construction Method, 190 6.2.2 Available Software, 192 6.2.3 Alpha Designs with Unequal Block Sizes, 192 Generalized Cyclic Incomplete Block Designs, 193 Designs Based on the Successive Diagonalizing Method, 194 6.4.1 Designs for t = Kk, 194 6.4.2 Designs with t = n2 , 194 Comparing Treatments with a Control, 195 6.5.1 Supplemented Balance, 196 6.5.2 Efficiencies and Optimality Criteria, 197 6.5.3 Balanced Treatment Incomplete Block Designs, 199 6.5.4 Partially Balanced Treatment Incomplete Block Designs, 205 6.5.5 Optimal Designs, 211 Row–Column Designs, 213 6.6.1 Introduction, 213 6.6.2 Model and Normal Equations, 213 6.6.3 Analysis of Variance, 215 6.6.4 An Example, 216 6.6.5 Regular Row–Column Designs, 230 6.6.6 Doubly Incomplete Row–Column Designs, 230 6.6.7 Properties of Row–Column Designs, 232 6.6.8 Construction, 237 6.6.9 Resolvable Row–Column Designs, 238

189

x

7

CONTENTS

Two-Level Factorial Designs 7.1 7.2

7.3

7.4

7.5 7.6

8

Introduction, 241 Case of Two Factors, 241 7.2.1 Definition of Main Effects and Interaction, 241 7.2.2 Orthogonal Contrasts, 243 7.2.3 Parameterizations of Treatment Responses, 244 7.2.4 Alternative Representation of Treatment Combinations, Main Effects, and Interaction, 247 Case of Three Factors, 248 7.3.1 Definition of Main Effects and Interactions, 249 7.3.2 Parameterization of Treatment Responses, 252 7.3.3 The x -Representation, 252 General Case, 253 7.4.1 Definition of Main Effects and Interactions, 254 7.4.2 Parameterization of Treatment Responses, 256 7.4.3 Generalized Interactions, 258 Interpretation of Effects and Interactions, 260 Analysis of Factorial Experiments, 262 7.6.1 Linear Models, 262 7.6.2 Yates Algorithm, 263 7.6.3 Variances of Estimators, 265 7.6.4 Analysis of Variance, 265 7.6.5 Numerical Examples, 267 7.6.6 Use of Only One Replicate, 278

Confounding in 2n Factorial Designs 8.1

8.2

8.3

8.4

241

Introduction, 279 8.1.1 A Simple Example, 279 8.1.2 Comparison of Information, 280 8.1.3 Basic Analysis, 280 Systems of Confounding, 283 8.2.1 Blocks of Size 2n−1 , 283 8.2.2 Blocks of Size 2n−2 , 283 8.2.3 General Case, 285 Composition of Blocks for a Particular System of Confounding, 289 8.3.1 Intrablock Subgroup, 289 8.3.2 Remaining Blocks, 290 Detecting a System of Confounding, 291

279

xi

CONTENTS

8.5 8.6

8.7 8.8 9

Using SAS for Constructing Systems of Confounding, 293 Analysis of Experiments with Confounding, 293 8.6.1 Estimation of Effects and Interactions, 293 8.6.2 Parameterization of Treatment Responses, 297 8.6.3 ANOVA Tables, 298 Interblock Information in Confounded Experiments, 303 Numerical Example Using SAS, 311

Partial Confounding in 2n Factorial Designs 9.1 9.2

9.3

9.4 9.5

9.6

9.7

9.8 9.9

9.10

Introduction, 312 Simple Case of Partial Confounding, 312 9.2.1 Basic Plan, 312 9.2.2 Analysis, 313 9.2.3 Use of Intra- and Interblock Information, 315 Partial Confounding as an Incomplete Block Design, 318 9.3.1 Two Models, 318 9.3.2 Normal Equations, 320 9.3.3 Block Contrasts, 322 Efficiency of Partial Confounding, 323 Partial Confounding in a 23 Experiment, 324 9.5.1 Blocks of Size 2, 324 9.5.2 Blocks of Size 4, 325 Partial Confounding in a 24 Experiment, 327 9.6.1 Blocks of Size 2, 327 9.6.2 Blocks of Size 4, 328 9.6.3 Blocks of Size 8, 328 General Case, 329 9.7.1 Intrablock Information, 330 9.7.2 The ANOVAs, 330 9.7.3 Interblock Information, 332 9.7.4 Combined Intra- and Interblock Information, 333 9.7.5 Estimation of Weights, 333 9.7.6 Efficiencies, 334 Double Confounding, 335 Confounding in Squares, 336 9.9.1 23 Factorial in Two 4 × 4 Squares, 337 9.9.2 24 Factorial in 8 × 8 Squares, 338 Numerical Examples Using SAS, 338 9.10.1 23 Factorial in Blocks of Size 2, 338 9.10.2 24 Factorial in Blocks of Size 4, 350

312

xii

CONTENTS

10 Designs with Factors at Three Levels 10.1 10.2

10.3 10.4 10.5

10.6

10.7

10.8

Introduction, 359 Definition of Main Effects and Interactions, 359 10.2.1 The 32 Case, 359 10.2.2 General Case, 363 Parameterization in Terms of Main Effects and Interactions, 365 Analysis of 3n Experiments, 366 Confounding in a 3n Factorial, 368 10.5.1 The 33 Experiment in Blocks of Size 3, 369 10.5.2 Using SAS PROC FACTEX, 370 10.5.3 General Case, 374 Useful Systems of Confounding, 374 10.6.1 Two Factors, 376 10.6.2 Three Factors, 376 10.6.3 Treatment Comparisons, 376 10.6.4 Four Factors, 379 10.6.5 Five Factors, 379 10.6.6 Double Confounding, 379 Analysis of Confounded 3n Factorials, 380 10.7.1 Intrablock Information, 381 10.7.2 The ANOVAs, 381 10.7.3 Tests of Hypotheses, 384 10.7.4 Interblock Information, 385 10.7.5 Combined Intra- and Interblock Information, 386 10.7.6 Estimation of Weights, 386 Numerical Example, 387 10.8.1 Intrablock Analysis, 387 10.8.2 Combined Analysis, 387

11 General Symmetrical Factorial Design 11.1 11.2 11.3 11.4 11.5 11.6

359

Introduction, 393 Representation of Effects and Interactions, 395 Generalized Interactions, 396 Systems of Confounding, 398 Intrablock Subgroup, 400 Enumerating Systems of Confounding, 402

393

xiii

CONTENTS

11.7

11.8 11.9 11.10

11.11

11.12

11.13

11.14

11.15

Fisher Plans, 403 11.7.1 Existence and Construction, 403 11.7.2 Identifying System of Confounding, 406 11.7.3 Application to Fisher Plans, 407 Symmetrical Factorials and Finite Geometries, 409 Parameterization of Treatment Responses, 410 Analysis of p n Factorial Experiments, 412 11.10.1 Intrablock Analysis, 413 11.10.2 Disconnected Resolved Incomplete Block Designs, 417 11.10.3 Analysis of Variance Tables, 420 Interblock Analysis, 421 11.11.1 Combining Interblock Information, 422 11.11.2 Estimating Confounded Interactions, 425 Combined Intra- and Interblock Information, 426 11.12.1 Combined Estimators, 427 11.12.2 Variance of Treatment Comparisons, 430 The s n Factorial, 431 11.13.1 Method of Galois Field Theory, 431 11.13.2 Systems of Confounding, 433 11.13.3 Method of Pseudofactors, 435 11.13.4 The (p1 × p2 × · · · × pm )n Factorial, 443 General Method of Confounding for the Symmetrical Factorial Experiment, 447 11.14.1 Factorial Calculus, 448 11.14.2 Orthogonal Factorial Structure (OFS), 452 11.14.3 Systems of Confounding with OFS, 453 11.14.4 Constructing Systems of Confounding, 457 11.14.5 Verifying Orthogonal Factorial Structure, 459 11.14.6 Identifying Confounded Interactions, 462 Choice of Initial Block, 463

12 Confounding in Asymmetrical Factorial Designs 12.1 12.2

Introduction, 466 Combining Symmetrical Systems of Confounding, 467 12.2.1 Construction of Blocks, 467 12.2.2 Properties of Kronecker Product Method, 469 12.2.3 Use of Pseudofactors, 474

466

xiv

CONTENTS

12.3

12.4

12.5

The GC/n Method, 477 12.3.1 Description of the Method, 478 12.3.2 Choice of Generators, 479 12.3.3 Loss of Degrees of Freedom, 479 Method of Finite Rings, 480 12.4.1 Mathematics of Ideals and Rings, 481 12.4.2 Treatment Representations, 483 12.4.3 Representation of Main Effects and Interactions, 483 12.4.4 Parameterization of Treatment Responses, 485 12.4.5 Characterization and Properties of the Parameterization, 488 12.4.6 Other Methods for Constructing Systems of Confounding, 491 Balanced Factorial Designs (BFD), 491 12.5.1 Definitions and Properties of BFDs, 493 12.5.2 EGD-PBIBs and BFDs, 499 12.5.3 Construction of BFDs, 502

13 Fractional Factorial Designs 13.1 13.2 13.3

13.4 13.5

13.6

Introduction, 507 Simple Example of Fractional Replication, 509 Fractional Replicates for 2n Factorial Designs, 513 13.3.1 The 21 Fraction, 513 13.3.2 Resolution of Fractional Factorials, 516 13.3.3 Word Length Pattern, 518 13.3.4 Criteria for Design Selection, 518 Fractional Replicates for 3n Factorial Designs, 524 General Case of Fractional Replication, 529 13.5.1 Symmetrical Factorials, 529 13.5.2 Asymmetrical Factorials, 529 13.5.3 General Considerations, 531 13.5.4 Maximum Resolution Design, 534 Characterization of Fractional Factorial Designs of Resolution III, IV, and V, 536 13.6.1 General Formulation, 536 13.6.2 Resolution III Designs, 538 13.6.3 Resolution IV Designs, 539 13.6.4 Foldover Designs, 543

507

xv

CONTENTS

13.7

13.8

13.9

13.6.5 Resolution V Designs, 546 Fractional Factorials and Combinatorial Arrays, 547 13.7.1 Orthogonal Arrays, 547 13.7.2 Balanced Arrays, 549 Blocking in Fractional Factorials, 549 13.8.1 General Idea, 549 13.8.2 Blocking in 2n− Designs, 550 13.8.3 Optimal Blocking, 551 Analysis of Unreplicated Factorials, 558 13.9.1 Half-Normal Plots, 558 13.9.2 Bar Chart, 562 13.9.3 Extension to Nonorthogonal Design, 563

14 Main Effect Plans 14.1 14.2

14.3

14.4

Introduction, 564 Orthogonal Resolution III Designs for Symmetrical Factorials, 564 14.2.1 Fisher Plans, 564 14.2.2 Collapsing Factor Levels, 567 14.2.3 Alias Structure, 567 14.2.4 Plackett–Burman Designs, 575 14.2.5 Other Methods, 577 Orthogonal Resolution III Designs for Asymmetrical Factorials, 582 14.3.1 Kronecker Product Design, 583 14.3.2 Orthogonality Condition, 583 14.3.3 Addelman–Kempthorne Methods, 586 Nonorthogonal Resolution III Designs, 594

15 Supersaturated Designs 15.1 15.2 15.3 15.4

15.5 15.6

564

Introduction and Rationale, 596 Random Balance Designs, 596 Definition and Properties of Supersaturated Designs, 597 Construction of Two-Level Supersaturated Designs, 598 15.4.1 Computer Search Designs, 598 15.4.2 Hadamard-Type Designs, 599 15.4.3 BIBD-Based Supersaturated Designs, 601 Three-Level Supersaturated Designs, 603 Analysis of Supersaturated Experiments, 604

596

xvi

CONTENTS

16 Search Designs 16.1 16.2 16.3

16.4

16.5

16.6

Introduction and Rationale, 608 Definition of Search Design, 608 Properties of Search Designs, 609 16.3.1 General Case, 609 16.3.2 Main Effect Plus One Plans, 611 16.3.3 Resolution V Plus One Plans, 614 16.3.4 Other Search Designs, 614 Listing of Search Designs, 615 16.4.1 Resolution III.1 Designs, 615 16.4.2 Resolution V.1 Designs, 616 Analysis of Search Experiments, 617 16.5.1 General Setup, 617 16.5.2 Noiseless Case, 618 16.5.3 Noisy Case, 625 Search Probabilities, 630

17 Robust-Design Experiments 17.1 17.2 17.3 17.4

17.5

18.2 18.3

633

Off-Line Quality Control, 633 Design and Noise Factors, 634 Measuring Loss, 635 Robust-Design Experiments, 636 17.4.1 Kronecker Product Arrays, 636 17.4.2 Single Arrays, 636 Modeling of Data, 638 17.5.1 Location and Dispersion Modeling, 638 17.5.2 Dual-Response Modeling, 641

18 Lattice Designs 18.1

608

Definition of Quasi-Factorial Designs, 649 18.1.1 An Example: The Design, 649 18.1.2 Analysis, 650 18.1.3 General Definition, 653 Types of Lattice Designs, 653 Construction of One-Restrictional Lattice Designs, 655 18.3.1 Two-Dimensional Lattices, 655 18.3.2 Three-Dimensional Lattices, 656 18.3.3 Higher-Dimensional Lattices, 657

649

xvii

CONTENTS

18.4

General Method of Analysis for One-Restrictional Lattice Designs, 657 18.5 Effects of Inaccuracies in the Weights, 661 18.6 Analysis of Lattice Designs as Randomized Complete Block Designs, 666 18.7 Lattice Designs as Partially Balanced Incomplete Block Designs, 669 18.8 Lattice Designs with Blocks of Size K , 670 18.9 Two-Restrictional Lattices, 671 18.9.1 Lattice Squares with K Prime, 671 18.9.2 Lattice Squares for General K, 675 18.10 Lattice Rectangles, 678 18.11 Rectangular Lattices, 679 18.11.1 Simple Rectangular Lattices, 680 18.11.2 Triple Rectangular Lattices, 681 18.12 Efficiency Factors, 682 19 Crossover Designs 19.1 19.2 19.3 19.4 19.5

19.6

19.7 19.8

Introduction, 684 Residual Effects, 685 The Model, 685 Properties of Crossover Designs, 687 Construction of Crossover Designs, 688 19.5.1 Balanced Designs for p = t, 688 19.5.2 Balanced Designs for p < t, 688 19.5.3 Partially Balanced Designs, 691 19.5.4 Strongly Balanced Designs for p = t + 1, 691 19.5.5 Strongly Balanced Designs for p < t, 692 19.5.6 Balanced Uniform Designs, 693 19.5.7 Strongly Balanced Uniform Designs, 693 19.5.8 Designs with Two Treatments, 693 Optimal Designs, 695 19.6.1 Information Matrices, 695 19.6.2 Optimality Results, 697 Analysis of Crossover Designs, 699 Comments on Other Models, 706 19.8.1 No Residual Effects, 706 19.8.2 No Period Effects, 707 19.8.3 Random Subject Effects, 707

684

xviii

CONTENTS

19.8.4 19.8.5 19.8.6 19.8.7

Autocorrelation Error Structure, 707 Second-Order Residual Effects, 708 Self and Mixed Carryover Effects, 710 Factorial Treatment Structure, 713

Appendix A Fields and Galois Fields

716

Appendix B

721

Finite Geometries

Appendix C Orthogonal and Balanced Arrays

724

Appendix D Selected Asymmetrical Balanced Factorial Designs

728

Appendix E

736

Exercises

References

749

Author Index

767

Subject Index

771

Preface

The project of revising Kempthorne’s 1952 book Design and Analysis of Experiments started many years ago. Our desire was to not only make minor changes to what had become a very successful book but to update it and incorporate new developments in the field of experimental design. Our involvement in teaching this topic to graduate students led us soon to the decision to separate the book into two volumes, one for instruction at the MS level and one for instruction and reference at the more advanced level. Volume 1 (Hinkelmann and Kempthorne, 1994) appeared as an Introduction to Experimental Design. It lays the philosophical foundation and discusses the principles of experimental design, going back to the ground-breaking work of the founders of this field, R. A. Fisher and Frank Yates. At the basis of this development lies the randomization theory as advocated by Fisher and the further development of these ideas by Kempthorne in the form of derived linear models. All the basic error control designs, such as completely randomized design, block designs, Latin square type designs, split-plot designs, and their associated analyses are discussed in this context. In doing so we draw a clear distinction among the three components of an experimental design: the error control design, the treatment design, and the sampling design. Volume 2 builds upon these foundations and provides more details about certain aspects of error control and treatment designs and the connections between them. Much of the effort is concentrated on the construction of incomplete block designs for various types of treatment structures, including “ordinary” treatments, control and test treatments, and factorial treatments. This involves, by necessity, a certain amount of combinatorics and leads, almost automatically, to the notions of balancedness, partial balancedness, orthogonality, and uniformity. These, of course, are also generally desirable properties of experimental designs and aspects of their analysis. In our discussion of ideas and methods we always emphasize the historical developments of and reasons for the introduction of certain designs. The development of designs was often dictated by computational aspects of the ensuing analysis, and this, in turn, led to the properties mentioned above. Even though xix

xx

PREFACE

in the age of powerful computers and the wide availability of statistical computer software these concerns no longer play the dominant feature, we remind the reader that such properties have general statistical appeal and often serve as starting points for new developments. Moreover, we caution the reader that not all software can be trusted all the time when it comes to the analysis of data from completely unstructured designs, apart from the fact that the interpretation of the results may become difficult and ambiguous. The development and introduction of new experimental designs in the last 50 years or so has been quite staggering, brought about, in large part, by an everwidening field of applications and also by the mathematical beauty and challenge that some of these designs present. Whereas many designs had their origin in agricultural field experiments, it is true now that these designs as well as modifications, extensions, and new developments were initiated by applications in almost all types of experimental research, including industrial and clinical research. It is for this reason that books have been written with special applications in mind. We, on the other hand, have tried to keep the discussion in this book as general as possible, so that the reader can get the general picture and then apply the results in whatever area of application is desired. Because of the overwhelming amount of material available in the literature, we had to make selections of what to include in this book and what to omit. Many special designs or designs for special cases (parameters) have been presented in the literature. We have concentrated, generally speaking, on the more general developments and results, providing and discussing methods of constructing rather large classes of designs. Here we have built upon the topics discussed in Kempthorne’s 1952 book and supplemented the material with more recent topics of theoretical and applications oriented interests. Overall, we have selected the material and chosen the depth of discussion of the various topics in order to achieve our objective for this book, namely to serve as a textbook at the advanced graduate level and as a reference book for workers in the field of experimental design. The reader should have a solid foundation in and appreciation of the principles and fundamental notions of experimental design as discussed, for example, in Volume 1. We realize that the material presented here is more than can be covered in a one-semester course. Therefore, the instructor will have to make choices of the topics to be discussed. In Chapters 1 through 6 we discuss incomplete block and row–column designs at various degrees of specificity. In Chapter 1 we lay the general foundation for the notion and analysis of incomplete block designs. This chapter is essential because its concepts permeate through almost every chapter of the book, in particular the ideas of intra- and interblock analyses. Chapters 2 through 5 are devoted to balanced and partially balanced incomplete block designs, their special features and methods of construction. In Chapter 6 we present some other types of incomplete block designs, such as α-designs and control-test treatment comparison designs. Further, we discuss various forms of row–column designs as examples of the use of additional blocking factors.

PREFACE

xxi

In Chapters 7 through 13 we give a general discussion of the most fundamental and important ideas of factorial designs, beginning with factors at two levels (Chapters 7 through 9), continuing with the case of factors with three levels (Chapter 10) through the general case of symmetrical and asymmetrical factorial designs (Chapters 11 and 12), and concluding with the important concept of fractional factorial designs (Chapter 13). In these chapters we return often to the notion of incomplete block designs as we discuss various systems of confounding of interaction effects with block effects. Additional topics involving factorial designs are taken up in Chapters 14 through 17. In Chapter 14 we discuss the important concept of main effect plans and their construction. This notion is then extended to supersaturated designs (Chapter 15) and incorporated in the ideas of search designs (Chapter 16) and robust-design or Taguchi experiments (Chapter 17). We continue with an extensive chapter about lattice designs (Chapter 18), where the notions of factorial and incomplete block designs are combined in a unique way. We conclude the book with a chapter on crossover designs (Chapter 19) as an example where the ideas of optimal incomplete row–column designs are complemented by the notion of carryover effects. In making a selection of topics for teaching purposes the instructor should keep in mind that we consider Chapters 1, 7, 8, 10, and 13 to be essential for the understanding of much of the material in the book. This material should then be supplemented by selected parts from the remaining chapters, thus providing the student with a good understanding of the methods of constructing various types of designs, the properties of the designs, and the analyses of experiments based on these designs. The reader will notice that some topics are discussed in more depth and detail than others. This is due to our desire to give the student a solid foundation in what we consider to be fundamental concepts. In today’s computer-oriented environment there exist a number of software programs that help in the construction and analysis of designs. We have chosen to use the Statistical Analysis System (SAS) for these purposes and have provided throughout the book examples of input statements and output using various procedures in SAS, both for constructing designs as well as analyzing data from experiments based on these designs. For the latter, we consider, throughout, various forms of the analysis of variance to be among the most important and informative tools. As we have mentioned earlier, Volume 2 is based on the concepts developed and described in Volume 1. Nevertheless, Volume 2 is essentially self-contained. We make occasional references to certain sections in Volume 1 in the form (I.xx.yy) simply to remind the reader about certain notions. We emphasize again that the entire development is framed within the context of randomization theory and its approximation by normal theory inference. It is with this fact in mind that we discuss some methods and ideas that are based on normal theory. There exist a number of books discussing the same types of topics that we exposit in this book, some dealing with only certain types of designs, but perhaps present more details than we do. For some details we refer to these books

xxii

PREFACE

in the text. A quite general, but less detailed discussion of various aspects of experimental design is provided by Cox and Reid (2000). Even though we have given careful attention to the selection of material for this book, we would be remiss if we did not mention that certain areas are completely missing. For example, the reader will not find any discussion of Bayesian experimental design. This is, in part, due to our philosophical attitude toward the Bayesian inferential approach (see Kempthorne, 1984; Hinkelmann, 2001). To explain, we strongly believe that design of experiment is a Bayesian experimentation process, not a Bayesian inference process, but one in which the experimenter approaches the experiment with some beliefs, to which he accommodates the design. It is interesting to speculate whether precise mathematical formulation of informal Bayesian thinking will be of aid in design. Another area that is missing is that of sequential design. Here again, we strongly believe and encourage the view that most experimentation is sequential in an operational sense. Results from one, perhaps exploratory, experiment will often lead to further, perhaps confirmatory, experimentation. This may be done informally or more formally in the context of sequential probability ratio tests, which we do not discuss explicitly. Thus, the selection and emphases are to a certain extent subjective and reflect our own interests as we have taught over the years parts of the material to our graduate students. As mentioned above, the writing of this book has extended over many years. This has advantages and disadvantages. My (K.H.) greatest regret, however, is that the book was not completed before the death of my co-author, teacher, and mentor, Oscar Kempthorne. I only hope that the final product would have met with his approval. This book could not have been completed without the help from others. First, we would like to thank our students at Virginia Tech, Iowa State University, and the University of Dortmund for their input and criticism after being exposed to some of the material. K.H. would like to thank the Departments of Statistics at Iowa State University and the University of Dortmund for inviting him to spend research leaves there and providing him with support and a congenial atmosphere to work. We are grateful to Michele Marini and Ayca Ozol-Godfrey for providing critical help with some computer work. Finally, we will never be able to fully express our gratitude to Linda Breeding for her excellent expert word-processing skills and her enormous patience in typing the manuscript, making changes after changes to satisfy our and the publisher’s needs. It was a monumental task and she did as well as anybody possibly could. Klaus Hinkelmann Blacksburg, VA May 2004

CHAPTER 1

General Incomplete Block Design

1.1 INTRODUCTION AND EXAMPLES One of the basic principles in experimental design is that of reduction of experimental error. We have seen (see Chapters I.9 and I.10) that this can be achieved quite often through the device of blocking. This leads to designs such as randomized complete block designs (Section I.9.2) or Latin square type designs (Chapter I.10). A further reduction can sometimes be achieved by using blocks that contain fewer experimental units than there are treatments. The problem we shall be discussing then in this and the following chapters is the comparison of a number of treatments using blocks the size of which is less than the number of treatments. Designs of this type are called incomplete block designs (see Section I.9.8). They can arise in various ways of which we shall give a few examples. In the case of field plot experiments, the size of the plot is usually, though by no means always, fairly well determined by experimental and agronomic techniques, and the experimenter usually aims toward a block size of less than 12 plots. If this arbitrary rule is accepted, and we wish to compare 100 varieties or crosses of inbred lines, which is not an uncommon situation in agronomy, we will not be able to accommodate all the varieties in one block. Instead, we might use, for example 10 blocks of 10 plots with different arrangements for each replicate (see Chapter 18). Quite often a block and consequently its size are determined entirely on biological or physical grounds, as, for example, a litter of mice, a pair of twins, an individual, or a car. In the case of a litter of mice it is reasonable to assume that animals from the same litter are more alike than animals from different litters. The litter size is, of course, restricted and so is, therefore, the block size. Moreover, if one were to use female mice only for a certain investigation, the block size would be even more restricted, say to four or five animals. Hence, Design and Analysis of Experiments. Volume 2: Advanced Experimental Design By Klaus Hinkelmann and Oscar Kempthorne ISBN 0-471-55177-5 Copyright  2005 John Wiley & Sons, Inc.

1

2

GENERAL INCOMPLETE BLOCK DESIGN

comparing more than this number of treatments would require some type of incomplete block design. Suppose we wish to compare seven treatments, T1 , T2 , T3 , T4 , T5 , T6 , T7 , say, using female mice, and suppose we have several litters with four females. We then could use the following incomplete block design, which, as will be explained later, is a balanced incomplete block design:

Litter

1

Animal 2 3

1 2 3 4 5 6 7

T1 T3 T7 T1 T2 T5 T2

T4 T6 T1 T2 T7 T3 T4

T7 T5 T2 T3 T3 T4 T5

4 T6 T7 T5 T6 T4 T1 T6

Notice that with this arrangement every treatment is replicated four times, and every pair of treatments occurs together twice in the same block; for example, T1 and T2 occur together in blocks 3 and 4. Many sociological and psychological studies have been done on twins because they are “alike” in many respects. If they constitute a block, then the block size is obviously two. A number of incomplete block designs are available for this type of situation, for example, Kempthorne (1953) and Zoellner and Kempthorne (1954). Blocks of size two arise also in some medical studies, when a patient is considered to be a block and his eyes or ears or legs are the experimental units. With regard to a car being a block, this may occur if we wish to compare brands of tires, using the wheels as the experimental units. In this case one may also wish to take the effect of position of the wheels into account. This then leads to an incomplete design with two-way elimination of heterogeneity (see Chapters 6 and I.10). These few examples should give the reader some idea why and how the need for incomplete block designs arises quite naturally in different types of research. For a given situation it will then be necessary to select the appropriate design from the catalogue of available designs. We shall discuss these different types of designs in more detail in the following chapters along with the appropriate analysis. Before doing so, however, it seems appropriate to trace the early history and development of incomplete block designs. This development has been a remarkable achievement, and the reader will undoubtedly realize throughout the next chapters that the concept of incomplete block designs is fundamental to the understanding of experimental design as it is known today.

GENERAL REMARKS ON THE ANALYSIS OF INCOMPLETE BLOCK DESIGNS

3

The origins of incomplete block designs go back to Yates (1936a) who introduced the concept of balanced incomplete block designs and their analysis utilizing both intra- and interblock information (Yates, 1940a). Other incomplete block designs were also proposed by Yates (1936b, 1937a, 1940b), who referred to these designs as quasi-factorial or lattice designs. Further contributions in the early history of incomplete block designs were made by Bose (1939, 1942) and Fisher (1940) concerning the structure and construction of balanced incomplete block designs. The notion of balanced incomplete block design was generalized to that of partially balanced incomplete block designs by Bose and Nair (1939), which encompass some of the lattice designs introduced earlier by Yates. Further extensions of the balanced incomplete block designs and lattice designs were made by Youden (1940) and Harshbarger (1947), respectively, by introducing balanced incomplete block designs for eliminating heterogeneity in two directions (generalizing the concept of the Latin square design) and rectangular lattices some of which are more general designs than partially balanced incomplete block designs. After this there has been a very rapid development in this area of experimental design, and we shall comment on many results more specifically in the following chapters.

1.2 GENERAL REMARKS ON THE ANALYSIS OF INCOMPLETE BLOCK DESIGNS The analysis of incomplete block designs is different from the analysis of complete block designs in that comparisons among treatment effects and comparisons among block effects are no longer orthogonal to each other (see Section I.7.3). This is referred to usually by simply saying that treatments and blocks are not orthogonal. This nonorthogonality leads to an analysis analogous to that of the two-way classification with unequal subclass numbers. However, this is only partly true and applies only to the analysis that has come to be known as the intrablock analysis. The name of the analysis is derived from the fact that contrasts in the treatment effects are estimated as linear combinations of comparisons of observations in the same block. In this way the block effects are eliminated and the estimates are functions of treatment effects and error (intrablock error) only. Coupled with the theory of least squares and the Gauss–Markov theorem (see I.4.16.2), this procedure will give rise to the best linear unbiased intrablock estimators for treatment comparisons. Historically, this has been the method first used for analyzing incomplete block designs (Yates, 1936a). We shall derive the intrablock analysis in Section 1.3. Based upon considerations of efficiency, Yates (1939) argued that the intrablock analysis ignores part of the information about treatment comparisons, namely that information contained in the comparison of block totals. This analysis has been called recovery of interblock information or interblock analysis.

4


Yates (1939, 1940a) showed for certain types of lattice designs and for the balanced incomplete block design how these two types of analyses can be combined to yield more efficient estimators of treatment comparisons. Nair (1944) extended these results to partially balanced incomplete block designs, and Rao (1947a) gave the analysis for any incomplete block design showing the similarity between the intrablock analysis and the combined intra- and interblock analysis. The intrablock analysis, as it is usually presented, is best understood by assuming that the block effects in the underlying linear model are fixed effects. But for the recovery of interblock information the block effects are then considered to be random effects. This leads sometimes to confusion with regard to the assumptions in the combined analysis, although it should be clear from the previous remark that then the block effects have to be considered random effects for both the intra- and interblock analysis. To emphasize it again, we can talk about intrablock analysis under the assumption of either fixed or random block effects. In the first case ordinary least squares (OLS) will lead to best linear unbiased estimators for treatment contrasts. This will, at least theoretically, not be true in the second case, which is the reason for considering the interblock information in the first place and using the Aitken equation (see I.4.16.2), which is also referred to as generalized (weighted ) least squares. We shall now derive the intrablock analysis (Section 1.3), the interblock analysis (Section 1.7), and the combined analysis (Section 1.8) for the general incomplete block design. Special cases will then be considered in the following chapters.

1.3 THE INTRABLOCK ANALYSIS 1.3.1 Notation and Model Suppose we have t treatments replicated r1 , r2 , . . . , rt times, respectively, and b blocks with k1 , k2 , . . . , kb units, respectively. We then have t i=1

ri =

b

kj = n

j =1

where n is the total number of observations. Following the derivation of a linear model for observations from a randomized complete block design (RCBD), using the assumption of additivity in the broad sense (see Sections I.9.2.2 and I.9.2.6), an appropriate linear model for observations from an incomplete block design is yij = µ + τi + βj + eij

(1.1)

(i = 1, 2, . . . , t; j = 1, 2, . . . , b; = 0, 1, . . . , nij ), where τi is the effect of the ith treatment, βj the effect of the j th block, and eij the error associated with the

5

THE INTRABLOCK ANALYSIS

observation yij . As usual, the eij contain both experimental and observational (sampling) error, that is, using notation established in Volume 1, eij = ij + ηij with ij representing experimental error and ηij representing observational error. Also, based on previous derivations (see I.6.3.4), we can treat the eij as i.i.d. random variables with mean zero and variance σe2 = σ2 + ση2 . Note that because nij , the elements of the incidence matrix N , may be zero, not all treatments occur in each block which is, of course, the definition of an incomplete block design. Model (1.1) can also be written in matrix notation as y = µI + Xτ τ + Xβ β + e

(1.2)

where I is a column vector consisting of n unity elements, Xβ is the observationblock incidence matrix   Ik1   Ik2   Xβ =   . ..   Ikb with Ikj denoting a column vector of kj unity elements (j = 1, 2, . . . , b) and Xτ = (x 1 , x 2 , . . . , x t ) is the observation-treatment incidence matrix, where x i is a column vector with ri unity elements and (n − ri ) zero elements such that x i x i = ri and x i x i = 0 for i = i (i, i = 1, 2, . . . , t). 1.3.2 Normal and Reduced Normal Equations The normal equations (NE) for µ, τi , and βj are then n µ+

t

ri τi +

b

j = G kj β

j =1

i=1

µ + ri τi + ri

b

j = Ti nij β

(i = 1, 2, . . . , t)

j =1

µ+ kj

t i=1

j = Bj nij τ i + kj β

(j = 1, 2, . . . , b)

(1.3)

6


where Ti =

yij = ith treatment total

j

Bj = G=

yij = j th block total

i

Ti =

i

Bj = overall total

j

Equations (1.3) can be written in matrix notation as 

In In

In Xτ

   X τ In   Xβ In

X τ X τ Xβ Xτ

    In y µ          = Xτ y   Xτ Xβ  τ         y Xβ Xβ X β β In Xβ

(1.4)

which, using the properties of I, Xτ , Xβ , can be written as  In In    RIt  KIb

Iτ R R N

     µ G           τ N ·  = T        B K β

Ib K

(1.5)

where R = diag (ri ) K = diag kj N = nij

t ×t b×b t ×b

(the incidence matrix)

T = (T1 , T2 , . . . , Tt ) B = (B1 , B2 , . . . , Bb ) τ = (τ1 , τ2 , . . . , τt ) β = (β1 , β2 , . . . , βb ) and the I’s are column vectors of unity elements with dimensions indicated by the subscripts. From the third set of equations in (1.5) we obtain µIb + τ) β = K −1 (B − N

(1.6)

7


Substituting (1.6) into the second set of (1.5), which can also be expressed µ + N β + R τ = T (since NIb = RIt ), leads to the reduced normal as N Ib equations (RNE) (see Section I.4.7.1) for τ τ = T − NK −1 B (R − N K −1 N )

(1.7)

Standard notation for (1.7) is

where

and

C τ =Q

(1.8)

C = R − N K −1 N

(1.9)

Q = T − N K −1 B

(1.10)

the (i, i ) element of C being cii = δii ri −

b nij ni j j =1

kj

with δii = 1 for i = i and = 0 otherwise, and the ith element of Q being Qi = Ti −

b nij Bj j =1

kj

And Qi is called the ith adjusted treatment total, the adjustment being due to the fact that the treatments do not occur the same number of times in the blocks. 1.3.3 The C Matrix and Estimable Functions We note that the matrix C of (1.9) is determined entirely by the specific design, that is, by the incidence matrix N . It is, therefore, referred to as the C matrix (sometimes also as the information matrix ) of that design. The C matrix is symmetric, and the elements in any row or any column of C add to zero, that is, CI = 0, which implies that r(C) = rank(C) ≤ t − 1. Therefore, C does not have an inverse and hence (1.8) cannot be solved uniquely. Instead we write a solution to (1.8) as τ = C−Q where C − is a generalized inverse for C (see Section 1.3.4).

(1.11)

8


If we write C = (c1 , c2 , . . . , ct ), where ci is the ith column of C, then the set of linear functions {ci τ , i = 1, 2, . . . , t} which span the totality of estimable functions of the treatment effects, has dimensionality r(C). Let c τ be an estimable function and c τ its estimator, with τ from (1.11). Then τ ) = E c C − Q E(c = c C − E(Q) = c C − Cτ τ to be an unbiased estimator for c τ for any τ , we then must have For c c C − C = c

(1.12)

Since CI = 0, it follows from (1.12) that c I = 0. Hence, only treatment contrasts are estimable. If r(C) = t − 1, then all treatment contrasts are estimable. In particular, all differences τi − τi (i = i ) are estimable, there being t − 1 linearly independent estimable functions of this type. Then the design is called a connected design (see also Section I.4.13.3). 1.3.4 Solving the Reduced Normal Equations In what follows we shall assume that the design is connected; that is, r(C) = t − 1. This means that C has t − 1 nonzero (positive) eigenvalues and one zero eigenvalue. From     1 1 1 1     C  ..  = 0 = 0  ..  . . 1 1 it follows then that (1, 1, . . . , 1) is an eigenvector corresponding to the zero eigenvalue. If we denote the nonzero eigenvalues of C by d1 , d2 , . . . , dt−1 and the corresponding eigenvectors by ξ 1 , ξ 2 , . . . , ξ t−1 with ξ i ξ i = 1 (i = 1, 2, . . . , t − 1) and ξ i ξ i = 0(i = i ), then we can write C in its spectral decomposition as C=

t−1 i=1

di ξ i ξ i

(1.13)

9


√ or with dt = 0 and ξ t = 1/ t(1, 1, . . . , 1), alternatively as C=

t

di ξ i ξ i

(1.14)

i=1

We note that ξ t ξ t = 1 and ξ i ξ t = 0 for i = 1, 2, . . . , t − 1. We now return to (1.8) and consider a solution to these equations of the form given by (1.11). Although there are many methods of finding generalized inverses, we shall consider here one particular method, which is most useful in connection with incomplete block designs, especially balanced and partially balanced incomplete block designs (see following chapters). This method is based on the following theorem, which is essentially due to Shah (1959). Theorem 1.1 Let C be a t × t matrix as given by (1.9) with r(C) = t − 1. = C + aII , where a = 0 is a real number, admits an inverse C −1 , and Then C −1 is a generalized inverse for C. C Proof as (a) We can rewrite C   1 1  = C + aII = C + a  C  ..  (1, 1, . . . , 1) = C + at ξ t ξ t . 1 and because of (1.13) = C

t−1

di ξ i ξ i + at ξ t ξ t

(1.15)

i=1

has nonzero roots d1 , d2 , . . . , dt−1 , dt = at and hence is nonClearly, C singular. Then −1 = C

t−1 1 1 ξ ξ ξ i ξ i + di at t t

(1.16)

i=1

−1 = C − we consider C C −1 C. From (1.13), (1.15), and (b) To show that C (1.16) we have C −1 = I = C

t i=1

ξ i ξ i

10


and −1 = CC

t−1 i=1

CC

−1

1 ξ i ξ i = I − ξ t ξ t = I − II t

(1.17)

C=C

which implies −1 = C − C We remark here already that determining C − for the designs in the following chapters will be based on (1.17) rather than on (1.14). −1 into (1.13) then yields a solution of the RNE (1.8); that is, Substituting C τ =C

−1

(1.18)

Q

We note that because of (1.8) and (1.16) −1 Q E ( τ) = E C =C

−1

−1

=C

E (Q) E (C τ)

−1

= C Cτ 1 = I − II τ t   τ1 − τ τ2 − τ    = .   ..  τt − τ with τ = 1/t i τi ; that is, E( τ ) is the same as if we had obtained a generalized inverse of C by imposing the condition i τi = 0. 1.3.5 Estimable Functions of Treatment Effects We know from the Gauss–Markov theorem (see Section I.4.16.2) that for any linear estimable function of the treatment effects, say c τ , τ ) = c τ E(c

(1.19)

is independent of the solution to the NE (see Section I.4.4.4). We have further var(c τ ) = c C

−1

c σe2

(1.20)

11


with a corresponding result (but same numerical value) for any other solution obtained, using an available software package (see Section 1.14). We shall elaborate on this point briefly. Let us rewrite model (1.2) as y = µI + X β β + X τ τ + e   µ = (I Xβ Xτ ) β  + e τ ≡ X + e with

X = (I : Xβ : X τ )

and

(1.21) (1.22)

= (µ, β , τ )

The NE for model (1.21) are X X∗ = X y

(1.23)

A solution to (1.23) is given by, say, ∗ = (X X)− X y for some (X X)− . Now (X X)− is a (1 + b + t) × (1 + b + t) matrix that we can partition conformably, using the form of X as given in (1.22), as   Aµµ Aµβ Aµτ (X X)− =  Aµβ Aββ Aβτ  (1.24) Aµτ Aβτ Aτ τ Here, Aτ τ is a t × t matrix that serves as the variance–covariance matrix for obtaining var(c τ ∗ ) = c Aτ τ c σe2

(1.25)

τ = c τ ∗ and also the numerical For any estimable function c τ we have c −1 values for (1.20) and (1.25) are the same. If we denote the (i, i ) element of C ii ii by c and the corresponding element of Aτ τ in (1.24) by a , then we have, for example, for c τ = τi − τi var ( τi − τi ) = cii − 2cii + ci i σe2 = a ii − 2a ii + a i i σe2 (1.26) For a numerical example and illustration of computational aspects, see Section 1.13.

12


1.3.6 Analyses of Variance It follows from general principles (see Section I.4.7.1) that the two forms of analysis of variance are as given in Tables 1.1 and 1.2. We shall henceforth refer to the analysis of variance in Table 1.1 as the treatment-after-block ANOVA or T | B-ANOVA as it is associated with the ordered model y = µI + Xβ β + Xτ τ + e whereas the analysis of variance in Table 1.2 is associated with the ordered model y = µI + X τ τ + Xβ β + e and hence shall be referred to as the block-after-treatment ANOVA or B | TANOVA. To indicate precisely the sources of variation and the associated sums of squares, we use the notation developed in Section I.4.7.2 for the general case as it applies to the special case of the linear model for the incomplete block Table 1.1 T|B-ANOVA for Incomplete Block Design Source

d.f.a

X β |I

b−1

X τ |I, X β

SS b B2 j j =1 t

t −1

kj

−

E(MS) G2 n

τi Qi

σe2 +

i=1

I |I, X β , X τ Total

n−b−t +1

ij a d.f.

σe2

Difference

n−1

yij2 −

τ Cτ t −1

G2 n

= degrees of freedom.

Table 1.2 B|T-ANOVA for Incomplete Block Design Source X τ |I

d.f. t −1

SS t T2 i

i=1

X β |I, Xτ I |I, X β , X τ Total

b−1 n−b−t +1 n−1

ri

−

G2 n

Difference From Table 1.1 ij

yij2 −

G2 n

INCOMPLETE DESIGNS WITH VARIABLE BLOCK SIZE

13

design, thereby avoiding the commonly used but not always clearly understood terms blocks ignoring treatments for (Xβ | I), treatments eliminating blocks for (X τ | I, Xβ ), and blocks eliminating treatments for (Xβ | I, Xτ ). The T | B-ANOVA follows naturally from the development of the RNE for the treatment effects. It is the appropriate ANOVA for the intrablock analysis as it allows to test the hypothesis H0 : τ1 = τ2 = · · · = τt by means of the (approximate) F test (see I.9.2.5) F =

SS(X τ | I, Xβ )/(t − 1) SS(I | I, Xβ , Xτ )/(n − b − t + 1)

(1.27)

Also MS(Error)= SS(I | I, X β , Xτ )/(n − b − t + 1) is an estimator for σe2 to be used for estimating var(c τ ) of (1.20). The usefulness of the B | T-ANOVA in Table 1.2 will become apparent when we discuss specific aspects of the combined intra- and interblock analysis in Section 1.10. At this point we just mention that SS(Xβ |, I, Xτ ) could have been obtained from the RNE for block effects. Computationally, however, it is more convenient to use the fact that SS(I | I, Xβ , Xτ ) = SS(I | I, Xτ , X β ) and then obtain SS(Xβ | I, X τ ) by subtraction. Details of computational procedures using SAS PROC GLM and SAS PROC Mixed (SAS1999–2000) will be described in Section 1.14. 1.4 INCOMPLETE DESIGNS WITH VARIABLE BLOCK SIZE In the previous section we discussed the intrablock analysis of the general incomplete block design; that is, a design with possibly variable block size and possibly variable number of replications. Although most designed experiments use blocks of equal size, k say, there exist, however, experimental situations where blocks of unequal size arise quite naturally. We shall distinguish between two reasons why this can happen and why caution may have to be exercised before the analysis as outlined in the previous section can be used: 1. As pointed out by Pearce (1964, p. 699): With much biological material there are natural units that can be used as blocks and they contain plots to a number not under the control of the experimenter. Thus, the number of animals in a litter or the number of blossoms in a truss probably vary only within close limits.

2. Although an experiment may have been set up using a proper design, that is, a design with equal block size, missing plots due to accidents during the course of investigation will leave one for purpose of analysis with a design of variable block size.

14


In both cases there are two alternatives to handle the situation. In case 1 one may wish to reduce all blocks to a constant size, thereby reducing the number of experimental units available. If experimental units are at a premium, this may not be the most desirable course of action. The other alternative is to use the natural blocks and then use the analysis as given in the previous section. Before doing so we mention that its validity will depend on one very important assumption, and that is the constancy of the variance σe2 for all blocks. In general, the size of σe2 will depend on the size of the blocks: The larger the blocks, the larger σe2 will be since it is in part a measure of the variability of the experimental units within blocks (see I.9.2.4). In fact, this is the reason for reducing the block size since it may also reduce the experimental error. Experience shows that such a reduction in σe2 is not appreciable for only modest reduction in block size. It is therefore quite reasonable to assume that σe2 is constant for blocks of different size if the number of experimental units varies only slightly. In case 2 one possibility is to estimate the missing values and then use the analysis for the proper design. Such a procedure, however, would only be approximate. The exact analysis then would require the analysis with variable block size as in case 1. Obviously, the assumption of constancy of experimental error is satisfied here if is was satisfied for the original proper design.

1.5 DISCONNECTED INCOMPLETE BLOCK DESIGNS In deriving the intrablock analysis of an incomplete block design in Section 1.3.4 we have made the assumption that the C matrix of (1.9) has maximal rank t − 1, that is, the corresponding design is a connected design. Although connectedness is a desirable property of a design and although most designs have this property, we shall encounter designs (see Chapter 8) that are constructed on purpose as disconnected designs. We shall therefore comment briefly on this class of designs. Following Bose (1947a) a treatment and a block are said to be associated if the treatment is contained in that block. Two treatments are said to be connected if it is possible to pass from one to the other by means of a chain consisting alternately of treatments and blocks such that any two adjacent members of the chain are associated. If this holds true for any two treatments, then the design is said to be connected, otherwise it is said to be disconnected (see Section I.4.13.3 for a more formal definition and Srivastava and Anderson, 1970). Whether a design is connected or disconnected can be checked easily by applying the definition given above to the incidence matrix N: If one can connect two nonzero elements of N by means of vertical and horizontal lines such that the vertices are at nonzero elements, then the two treatments are connected. In order to check whether a design is connected, it is sufficient to check whether a given treatment is connected to all the other t − 1 treatments. If a design is disconnected, it follows then that (possibly after suitable relabeling of the treatments) the matrix NN and hence C consist of disjoint block diagonal matrices such that the treatments associated with one of these submatrices are connected with each other.

15

DISCONNECTED INCOMPLETE BLOCK DESIGNS

Suppose C has m submatrices, that is,    C= 



C1

   

C2 ..

. Cm

m where C ν is tν × tν ν=1 tν = t . It then follows that rank (C ν ) = tν − 1(ν = 1, 2, . . . , m) and hence rank(C) = t − m. The RNE is still of the form (1.8) −1 we now have, modifying with a solution given by (1.11), where in C − = C Theorem 1.1,   a1 II   a2 II  =C+ C   ..   . am II

Table 1.3 T|B-ANOVA for Disconnected Incomplete Block Design Source

d.f.

X β |I

b−1

SS Bj2 j

X τ |I, X β

t −m

kj

G2 n

−

τi Qi

i

I |I, X β , X τ

n−b−t +m n−1

Total

Difference

yij2 −

ij

Table 1.4

B|T-ANOVA for Disconnected Incomplete Block Design

Source Xτ |I

d.f. t −1

SS T2 i

i

Xβ |I, X τ I |I, Xτ , Xβ Total

G2 n

b−m n−t −b+m n−1

ri

−

G2 n

Difference From Table 1.3 ij

yij2 −

G2 n

16


with aν (ν = 1, 2, . . . , m) arbitrary constants (= 0) and the II matrices are of appropriate dimensions. Following the development in Section 1.3.6, this then leads to the ANOVA tables as given in Tables 1.3 and 1.4.

1.6 RANDOMIZATION ANALYSIS So far we have derived the analysis of data from incomplete block designs using a Gauss–Markov linear model as specified in (1.1). We have justified the appropriate use of such an infinite population theory model in our earlier discussions of error control designs (see, e.g., Sections I.6.3 and I.9.2) as a substitute for a derived, that is, finite, population theory model that takes aspects of randomization into account. In this section we shall describe in mathematical terms the randomization procedure for an incomplete block design, derive an appropriate linear model, and apply it to the analysis of variance. This will show again, as we have argued in Section I.9.2 for the RCBD, that treatment effects and block effects cannot be considered symmetrically for purposes of statistical inference. 1.6.1 Derived Linear Model Following Folks and Kempthorne (1960) we shall confine ourselves to proper (i.e., all kj = k), equireplicate (i.e., all ri = r) designs. The general situation is then as follows: We are given a set of b blocks, each of constant size k ; a master plan specifies b sets of k treatments; these sets are assigned at random to the blocks; in each block the treatments are assigned at random to the experimental units (EU). This randomization procedure is described more formally by the following design random variables: αju =

1 0

if the uth set is assigned to the j th block otherwise

(1.28)

and δjuv

=

  1 

0

if the uv treatment is assigned to the th unit of the j th block otherwise

(1.29)

The uv treatment is one of the t treatments that, for a given design, has been assigned to the uth set. Assuming additivity in the strict sense (see Section I.6.3), the conceptual response of the uv treatment assigned to the th EU in the j th block can be written as Tj uv = Uj + Tuv (1.30)

17

RANDOMIZATION ANALYSIS

where Uj is the contribution from the th EU in the j th block and Tuv is the contribution from treatment uv. We then write further Tj uv = U .. + (U j. − U .. ) + (Uj − U j. ) + T .. + (Tuv − T .. ) = µ + bj + τuv + uj

(1.31)

where µ = U .. + T .. is the overall mean bj = U j. − U .. is the effect of the j th block (j = 1, 2, . . . , b) τuv = tuv − T .. is the effect of the uv treatment (u = 1, 2, . . . , b; v = 1, 2, . . . , k) uj = Uj − U j. is the unit error ( = 1, 2, . . . , k) with j bj = 0 = uv τuv = uj . We then express the observed response for the uv treatment, yuv , as yuv =

j

αju δjuv Tj uv

= µ + τuv +

αju bj +

j

j

αju δjuv uj

= µ + τuv + βu + ωuv where βu =

(1.32)

αju bj

(1.33)

j

is a random variable with E(βu ) = 0

E(βu2 ) =

1 2 bj b

E(βu βu ) =

j

Also, ωuv =

1 bj2 b(b − 1)

(u = u )

j

j

αju δjuv uj

(1.34)

18


is a random variable with E(ωuv ) = 0

1 2 uj bk

2 E(ωuv )=

j

E(ωuv ωuv ) = −

1 u2j bk(k − 1) j

E(ωuv ωu v ) = 0

(v = v )

(u = u )

In deriving the properties of the random variables βu and ωuv we have used, of course, the familiar distributional properties of the design random variables αju and δjuv , such as 1 P (αju = 1) = b P (αju = 1 | αju = 1) = 0 (j = j )

1 b(b − 1) 1 = 1) = k = 1) = 0

(u = u , j = j )

P (αju = 1 | αju = 1) = P (δjuv P (δjuv = 1) | (δjuv

( = )

1 k(k − 1) 1 = 1) = 2 k

( = , v = v )

P (δjuv = 1) | (δjuv = 1) =

P (δjuv = 1) | (δju v

(j = j , u = u )

and so on.

1.6.2 Randomization Analysis of ANOVA Tables Using model (1.32) and its distributional properties as induced by the design random variables αju and δjuv , we shall now derive expected values of the sums of squares in the analyses of variance as given in Tables 1.1 and 1.2: 1. E(SS Total) = E

(yuv − y .. )2

uv

=E =

uv

(τuv + βu + ωuv )2

uv 2 τuv +k

j

bj2 +

j

u2j

19


2. E[SS(X β | I)] = E

(y u. − y .. )2

uv

= kE =k

(τ u. + βu +

u

τ 2u. + k

u

1 ωuv )2 k v

bj2

j

n−t −b+1 2 3. E[SS (I | I, X β , Xτ )] = uj b(k − 1) j

since the incomplete block designs considered are unbiased. 4. E[SS(X τ | I, Xβ )] can be obtained by subtraction. 5. To obtain E[SS(X τ | I)] let

w γuv =

γuw =

  1  1 0

0

if the wth treatment corresponds to the uv index (w = 1, 2, . . . , t) otherwise if the wth treatment occurs in the uth block otherwise

where

w γuv = γuw

ν

and

γuw = r

u

Then 

2  1 1 w w γuv yuv − γuv yuv  E[SS(X τ | I)] = E  r w t uv w uv 

2  1 w =E rτw + γuw βu + γuv ωuv  r w u uv

=r

w

2 2 1 1 w τw2 + E γuw βu + E γuv ωuv r w r u w uv

20


=r

w

  γuw βu2 + γuw γuw βu βu 1  τw2 + E u uu r w u=u

 w 2 w w γuv ωuv + γuv γuv ωuv ωuv 1  + E uv u vv r w u=v

+

Now E  E

uu vv u=u

=r

γuw βu2

u

γuw γuw βu βu

uu

w w γuv γu v ωuv ωu v

j

 =−

γuw γuw

uu u=u

=−

1 bj2 b(b − 1) j

γuw (r − γuw )

u

=−

E  E

w 2 γuv ωuv = r

w w w γuv γuv ωuv ωuv

νν ν=ν

u

 E

w w γuv γu v ωuv ωu v

uu u=u

1 bj2 b(b − 1) j

r(r − 1) 2 bj b(b − 1) j

uv



1 2 bj b

u=u



1 2 uu bk j



w w  = 0 since γuv γuv

= 0

 =0

and hence E[(Xτ | I] = r

w

τw2 +

t (b − r) 2 t 2 bj + uj b(b − 1) bk j

j

Thus, we have for the mean squares (MS) from Tables 1.1 and 1.2 the expected values under randomization theory as given in Tables 1.5 and 1.6, respectively.

21


Table 1.5

E(MS) for T|B-ANOVA

Source

E(MS)

Xβ | I X τ | I, Xβ

k 2 k 2 τ u. + bj b−1 u b−1 j 1 2 2 2 (t − 1) uj + τuv − k τ u. b(k − 1) uv u j

I | I, Xβ , X τ

1 u2j b(k − 1) j

Table 1.6 E(MS) for B|T-ANOVA Source

E(MS)

Xτ | I

t t (b − r) r 2 u2j + bj2 + τ bk(t − 1) b(b − 1)(t − 1) t −1 w w

X β | I, X τ

bk − t 2 t −k u2j + bj b(b − 1)k(k − 1) (b − 1)2

I | I, Xβ , Xτ

1 u2j b(k − 1)

j

j

j

j

j

If we define

1 u2j = σu2 b(k − 1) j

and

1 2 bj = σβ2 b−1 j

we can then express the expected values for the three important mean squares in ANOVA Tables 1.5 and 1.6 as 2 2 ! uv τuv − k u τ u. 2 (1.35) E MS(X τ | I, Xβ ) = σu + t −1 ! E MS(X β | I, Xτ ) =

t −k n−t 2 σu2 + σ (b − 1)k b−1 β

! E MS(I | I, Xβ , Xτ ) = σu2

(1.36) (1.37)

22


We make the following observations: 1. The quadratic form in the τuv in (1.35) is just a different way of writing τ Cτ in Table 1.1. Both expressions indicate that the quadratic form depends on the particular design chosen, and both equal zero when all the treatment effects are the same. 2. It follows from (1.35) and (1.37) that, based on the equality of the E(MS) under H0 : τ1 = τ2 = · · · = τt , the ratio MS(X τ | I, Xβ )/MS(I | I, Xβ , X τ )

3.

4. 5.

6.

(1.38)

provides a test criterion for testing the above hypothesis. In fact, Ogawa (1974) has shown that the asymptotic randomization distribution of (1.38) is an F distribution with t − 1 and n − t − b + 1 d.f. We interpret this again to mean that the usual F test is an approximation to the randomization test based on (1.38). Considering (1.36) and (1.37), there does not exist an exact test for testing the equality of block effects. This is in agreement with our discussion in Section I.9.2 concerning the asymmetry of treatment and block effects. For k = t and r = b, that is, for the RCBD, the results of Tables 1.5 and 1.6 agree with those in Table 9.1 of Section I.9.2. With treatment-unit additivity in the broad sense (see Section I.6.3.3) the expressions in (1.35), (1.36), and (1.37) are changed by adding σν2 + ση2 to the right-hand sides (recall that σu2 + σν2 + ση2 ≡ σ2 + ση2 ≡ σe2 ). Remarks (2) and (3) above remain unchanged. For the recovery of interblock information (to be discussed in Section 1.7), we need to estimate σβ2 (or a function of σβ2 ). Clearly, under the assumption of additivity in the broad sense, this cannot be done considering that ! E MS Xβ | I, X τ = σν2 + ση2 +

t −k n−t 2 σ2 + σ (b − 1)k u b − 1 β

It is for this reason only that we shall resort to the approximation ! n−t 2 n−t 2 σβ = σe2 + σ E MS Xβ | I, Xτ ≈ σν2 + ση2 + σu2 + b−1 b−1 β (1.39) which is the expected value based on an infinite population theory model [see (1.49) and (1.50)]. For a different approach to randomization analysis, see Calinski and Kageyama (2000).

INTERBLOCK INFORMATION IN AN INCOMPLETE BLOCK DESIGN

23

1.7 INTERBLOCK INFORMATION IN AN INCOMPLETE BLOCK DESIGN 1.7.1 Introduction and Rationale As mentioned earlier, Yates (1939, 1940a) has argued that for incomplete block designs comparisons among block totals (or averages) contain some information about treatment comparisons, and he referred to this as recovery of interblock information. The basic idea is as follows. Consider, for purposes of illustration, the following two blocks and their observations from some design: Block 1:

y51 ,

y31 ,

y11

Block 2:

y22 ,

y42 ,

y32

Let B1 = y51 + y31 + y11 and B2 = y22 + y42 + y32 represent the block totals. Using model (1.1) we can write B1 − B2 = (τ5 + τ2 + τ1 ) − (τ2 + τ4 + τ3 ) + 3β1 − 3β2 + (e51 + e31 + e11 ) − (e22 + e42 + e32 ) Assuming now that the βj are random effects with mean zero, we find E(B1 − B2 ) = τ1 + τ5 − τ2 − τ4 It is in this sense that block comparisons contain information about treatment comparisons. We shall now formalize this procedure. 1.7.2 Interblock Normal Equations Consider the model equation (1.1) y = µI + Xτ τ + Xβ β + e where β is now assumed to be a random vector with E(β) = 0 and var(β) = σβ2 I . As pointed out above the interblock analysis is based on block totals rather

24


than on individual observations, that is, we now consider   k1  k2    Xβ y =  ..  µ + N τ + Kβ + X β e . kb

(1.40)

  k1  k2    E(Xβ y) =  ..  µ + N τ .

We then have

(1.41)

kb and the variance–covariance matrix under what we might call a double error structure with both β and e being random vectors var(X β y) = K 2 σβ2 + Kσe2 =K I+ = Lσe2

σβ2 σe2

K σe2

(say)

and "

L = diag{j } = diag kj 1 +

σβ2 σe2

# kj

with w= and wj = and ρj =

σe2

" = diag

w kj wj

# = diag{kj ρj }

1 σe2

(1.42)

1 + kj σβ2

(1.43)

σe2 + kj σβ2 w = wj σe2

(1.44)

The quantities w and wj of (1.42) and (1.43) are referred to as intrablock and interblock weights, respectively, as w is the reciprocal of the intrablock variance,

INTERBLOCK INFORMATION IN AN INCOMPLETE BLOCK DESIGN

25

σe2 , and wj is the reciprocal of the interblock variance, that is, var(Bj ) on a per observation basis, or var(Bj /kj ) = σe2 + kj σβ2 . We then use as our “observation” vector z = L−1/2 Xβ y (1.45) which has var(z) = I σe2 and hence satisfies the Gauss–Markov conditions. If we write (1.41) as µ E(Xβ y) = (k N ) τ with k = (k1 , k2 , . . . , kb ) , we have from (1.45) that −1/2

E(z) = L

µ (k N ) τ

The resulting NE, which we shall refer to as the interblock NE, is then given by ∗ k k −1 µ L (k N ) ∗ = L−1 Xβ y N τ N

(1.46)

or explicitly as 

kj2

  j j    kj  n1j  j  j  ..   .   kj  ntj j j

j

kj n1j j

n21j j

j

.. . n1j ntj j j

   kj kj ··· ntj Bj   j      j j  j  µ∗   n1 j  n1j ntj    ∗   B j   τ1   ···  j j      j  .  =  j    ..   .  .. ..     . ∗    τt   2 n j t ntj   Bj   ··· j j j

(1.47)

j

It can be seen easily that the rank of the coefficient matrix in (1.47) is t. To solve the interblock NE, we take µ∗ = 0 and hence reduce the set to the following t

26


equations in τ1∗ , τ2∗ , . . . , τt∗ where we have used the fact that j = kj ρj : 

n21j

ρj−1

 kj  j   ρj−1   n n 2j 1j  kj   .  ..     ρj−1 ntj n1j kj

n1j n2j

j

n22j

ρj−1 kj

ρj−1 kj

.. .

ntj n2j

ρj−1 kj

···

ρj−1



 kj     τ∗  1 ρj−1     τ ∗  ··· n2j ntj  2 kj   .     ..   ..  .  ∗  τt  ρj−1  2 ··· ntj kj n1j ntj

j

  ρj−1 n1j Bj    j kj       ρ −1    j  n B 2j j    =  j kj    ..     .      ρ −1  j  ntj Bj  kj

(1.48)

j

The solution to the equations (1.48) is referred to as the interblock information about the treatment effects, with E(τi∗ ) = µ + τi + const. ·

t

(µ + τi )

i =1

Hence E

i

ci τi∗

=

c i τi

for

ci = 0

i

We note here that typically (see Kempthorne, 1952) the interblock analysis is derived not in terms of the “observations” z [as given in (1.45)] but rather in terms of the block totals Xβ y. The resulting NE are then simply obtained by using L = I in (1.46) and subsequent equations. The reason why we prefer our description is the fact that then the intra- and interblock information can be combined additively to obtain the so-called combined analysis (see Section 1.8) rather than in the form of a weighted average (see Kempthorne, 1952).

27

COMBINED INTRA- AND INTERBLOCK ANALYSIS

1.7.3 Nonavailability of Interblock Information We conclude this section with the following obvious remarks: 1. For the special case kj = t for all j, ri = b for all i and all nij = 1, we have, of course, the RCBD. Then the elements in the coefficient matrix of (1.48) are all identical, and so are the right-hand sides. Consequently, (1.48) reduces to a single equation b τi∗ = Bj i

j

and no contrasts among the τi are estimable. Expressed differently, any contrast among block totals estimates zero, that is, no interblock information is available. 2. For a design with b < t (and such incomplete block designs exist as we shall see in Chapter 4; see also the example in Section 1.7.1), the rank of the coefficient matrix of (1.48), N L−1 N , is less than t. Hence not all µ + τi are estimable using block totals, which means that interblock information is not available for all treatment contrasts. 1.8 COMBINED INTRA- AND INTERBLOCK ANALYSIS 1.8.1 Combining Intra- and Interblock Information The two types of information about treatment effects that we derived in Sections 1.3.2 and 1.7.2 can be combined to yield the “best” information about estimable functions of treatment effects. All we need to do is to add the coefficient matrices from the intrablock RNE (1.8) and the interblock NE (1.48) and do the same for the corresponding right-hand sides. This will lead to a system of equations in τ1∗∗ , τ2∗∗ , . . . , τt∗∗ , say, and the solution to these equations will lead to the combined intra- and interblock estimators for treatment contrasts. In the following section we shall derive the equations mentioned above more directly using the method of generalized least squares, that is, by using the Aitken equations described in Section I.4.16.2 1.8.2 Linear Model In order to exhibit the double error structure that characterizes the underlying assumptions for the combined analysis, we rewrite model (1.1) as yj = µ + τj + βj + ej

(1.49)

where j = 1, 2, . . . , b; = 1, 2, . . . , kj ; τj denotes the effect of the treatment applied to the th experimental unit in the j th block, the βj are assumed to be i.i.d. random variables with E(βj ) = 0, var(βj ) = σβ2 , and the ej are i.i.d.

28


random variables with E(ej ) = 0, var(ej ) = σe2 . We then have E(yj ) = µ + τj and

 2 σβ + σe2    cov yj , yj = σβ2    0

(1.50)

for j = j , = for j = j , =

(1.51)

otherwise

To use matrix notation it is useful to arrange the observations according to blocks, that is, write the observation vector as y = y11 , y12 , . . . , y1k1 , y21 , y22 , . . . , yb1 , yb2 , . . . , ybkb Letting X = (I Xτ ) we rewrite (1.50) as

µ E(y) = X τ

and the variance–covariance (1.51) as  V1  V2  var(y) =   0

(1.52) 

..

.

0  2  σe ≡ V σe2  Vb

(1.53)

where V j is given by V j = I kj +

σβ2 σe2

Ikj Ikj

(1.54)

1.8.3 Normal Equations Applying now the principles of least squares to the model (1.52) with covariance structure (1.53) yields the Aitken equations (see Section I.4.16): µ −1 (X V X) (1.55) = X V −1 y τ where −1 −1 V −1 = diag V −1 , V , . . . , V b 1 2

29

COMBINED INTRA- AND INTERBLOCK ANALYSIS

and V −1 j = I kj − With

σβ2

1 =w σe2

σe2

and

Ikj Ikj

σe2 + kj σβ2

wj w

(1.56)

1 = wj + kj σβ2

= ρj−1

Eq. (1.56) can be written as V −1 j = I kj −

1 − ρj−1 kj

Ikj Ikj

and hence V

−1

= I n − diag

1 − ρ1−1 Ik1 Ik1 , k1

1 − ρ2−1 1 − ρb−1 Ik2 Ik2 , . . . , Ikb Ikb k2 kb

(1.57)

Further, if we let 1 − pj−1 /kj = δj (j = 1, 2, . . . , b), we have

X V −1

 1 − ρ1−1 Ik1    δ n I  1 11 k1 =X −   ···  δ1 nt1 Ik1

1 − ρ2−1 Ik2

···

δ2 n12 Ik2

···

δb n1b Ikb

···

···

···

δ2 nt2 Ik2

···

δb ntb Ikb

1 − ρb−1 Ikb

       · · · 

and X V −1 X 

kj ρj−1

  n1j ρj−1  =   · · · ntj ρj−1

n1j ρj−1 r1 − δj n21j ··· − δj n1j ntj

− −

n2j ρj−1

···

δj n1j n2j ···

··· − ···

δj n2j ntj

···

ntj ρj−1



  δj n1j ntj    ···  2 rt − δj ntj

(1.58)

30




 ρj−1 Bj     T − δj n1j Bj   1    X V−1 y =   ..     .     δj Ntj Bj Tt −

(1.59)

By inspection one can verify that in the coefficient matrix (1.58) the elements in rows 2 to t + 1 add up to the elements in row 1, which shows that (1.55) is not of full rank; in fact, it is of rank t. The easiest way to solve these equations then is to impose the condition µ = 0. This means that we eliminate the first row and first column from (1.58) and the first element in (1.59) and solve the resulting system of t equations in the t unknowns τ 1, τ 2, . . . , τ t . If we define S = diag(ρj ), then this system of equations resulting from (1.58) and (1.59) can be written as % & R − N K −1 I − S −1 τ = T − N K −1 I − S −1 B (1.60) which we write for short as A τ =P

(1.61)

with A and P as described above. The solution then is τ = A−1 P

(1.62)

or, in terms of a generalized inverse for the original set of NE (1.55) ( ' ( ' 0 0 µ X V −1 y = 0 A−1 τ

(1.63)

with E τ i = µ + τi

(i = 1, 2, . . . , t)

If we denote the (i, i ) element of A−1 by a ii , then τ i = a ii + a i i − 2a ii σe2 var τ τi − i − τi = var

(1.64)

More generally, the treatment contrast c τ is estimated by c τ with variance τ = c A−1 cσe2 (1.65) var c

RELATIONSHIPS AMONG INTRABLOCK, INTERBLOCK, AND COMBINED ESTIMATION

31

Expression (1.65) looks deceptively simple, but the reader should keep in mind that the elements of A−1 depend on σβ2 and σe2 . We shall return to estimating (1.65) in Section 1.10. Finally, we note that the equations (1.60) show a striking similarity to the intrablock NE (1.7), except that the system (1.60) is of full rank and the elements of its coefficient matrix depend on the unknown parameters σβ2 and σe2 . 1.8.4 Some Special Cases As a special case of the above derivations we mention explicitly the equireplicate, proper design, that is, the design with all ri = r and all kj = k. We then define w =

1 σe2 + kσβ2

(1.66)

w w

(1.67)

and ρ= and write (1.60) as ( ' 1 1 −1 1−ρ 1 − ρ −1 N B NN τ =T − rI − k k

(1.68)

We shall comment briefly on the set of equations (1.68) for two special cases: 1. If ρ −1 = 0, that is, σβ2 = ∞, then (1.68) reduces to the NE (1.7) for the intrablock analysis. This means, of course, that in this case no interblock information is available. This is entirely plausible and suggests further that for “large” σβ2 the interblock information is very weak and perhaps not worthwhile considering. 2. If ρ −1 = 1, that is, σβ2 = 0, the solution to (1.68) is the same as that obtained for the completely randomized design (CRD) with the restriction µˆ = 0. This, of course, is a formal statement and should not imply that in this case the observations should be analyzed as if a CRD had been used.

1.9 RELATIONSHIPS AMONG INTRABLOCK, INTERBLOCK, AND COMBINED ESTIMATION On several occasions we have pointed out that there exist certain relationships among the different types of analysis for incomplete block designs. It is worthwhile to exposit this in a little detail.

32


1.9.1 General Case For the full model y = In µ + X τ τ + Xβ β + e with the double error structure we have E(y) = Iµ + Xτ τ and var(y) = X β Xβ σβ2 + I σe2 = (I + γ X β X β )σe2 ≡ V σe2 as in (1.53) with γ = σβ2 /σe2 . Then, as explained in Section 1.8, the estimators of estimable functions are obtained from the Aitken equations by minimizing (y − Iµ − X τ τ ) V −1 (y − Iµ − Xτ τ )

(1.69)

with respect to µ and τ . To simplify the algebra, let = Iµ + Xτ τ and then write ψ = y − . Expression (1.69) can now be written simply as ψ V −1 ψ. We then write ψ = P ψ + (I − P )ψ where, again for brevity, we write P = P x β = Xβ (Xβ Xβ )−1 Xβ = Xβ K −1 Xβ

(1.70)

Then ψ V −1 ψ = [ψ P + ψ (I − P )]V −1 [P ψ + (I − P )ψ]

(1.71)

In order to expand the right-hand side of (1.71) we make use of the following results. From (I − P )V = (I − P )(I + γ X β Xβ ) = (I − P ) [using (1.70)] it follows that (I − P )V −1 = (I − P )


33

and hence (I − P )V −1 P = (I − P )P = 0 Thus, (1.71) reduces to ψ V −1 ψ = ψ P V −1 P ψ + ψ (I − P )ψ

(1.72)

To handle the term ψ P V −1 P ψ we note that V −1 in (1.57) can be rewritten as V −1 = I − γ X β (I + γ X β Xβ )−1 Xβ so that P V −1 P = P − γ X β (I + γ X β Xβ )−1 Xβ % & = Xβ K −1 − γ (I + γ K)−1 Xβ '

( 1 = Xβ diag Xβ kj (1 + γ kj ) = Xβ (K + γ K 2 )−1 X β Hence −1 K + γ K2 X β ψ ψ P V −1 P ψ = X β ψ

(1.73)

Then we note that Xβ ψ is the vector of block totals of the vector ψ. Since ψ = y − , we have var Xβ ψ = var X β y 2 = Xβ X β σβ2 + X β Xβ σe2 = σe2 K + γ K 2 Hence ψ P V −1 P ψ is equal to −1 Q2 ≡ [B − E(B)] K + γ K 2 [B − E(B)] with B = Xβ y being the vector of block totals.

(1.74)

34


The second expression in (1.72) is actually the quadratic form that needs to be minimized to obtain the RNE for (see Section I.4.7.1), and this is equivalent to minimizing Q1 ≡ (y − Xτ τ ) (I − P )(y − Xτ τ )

(1.75)

We thus summarize: To obtain the intrablock estimator, we minimize Q1 ; to obtain the interblock estimator, we minimize Q2 ; and to obtain the combined estimator, we minimize Q1 + Q2 . 1.9.2 Case of Proper, Equireplicate Designs We now consider the case with R = rI and K = kI . The intrablock NE are [see (1.7)]

1 rI − N N τˆ = Q k

For the interblock observational equations B = kIµ + N τ + error or, absorbing µ into τ , that is, replacing Iµ + τ by τ , B = N τ + error we have the interblock NE [see (1.47)] N N τ ∗ = NB As we have pointed out earlier (see Section 1.8) and as is obvious from (1.72), the combined NE are ( ' 1 1 1 NN NB τ =Q+ rI − N N + k k(1 + γ k) k(1 + γ k)

(1.76)

The form of (1.76) shows again two things: 1. The intrablock and interblock NE are related. 2. The matrix N N determines the nature of the estimators, both intrablock and interblock.


35

The properties of N N can be exploited to make some statements about estimable functions of treatment effects. Being real symmetric, NN is orthogonally diagonalizable. We know that 1 rI − NN I = 0 k or N N I = rkI √ so that one root of NN is rk ≡ δt , say, with associated eigenvector (1/ t) I ≡ ξ t , say. Suppose ξ 1 , ξ 2 , . . . , ξ t−1 complete the full set of orthonormal eigenvectors with associated eigenvalues δ1 , δ2 , . . . , δt−1 . Then with O = (ξ 1 , ξ 2 , . . . , ξ t ) and NN ξ i = δi ξ i the NE are equivalent to O

1 τ = O Q rI − N N OO k

(1.77)

O N N OO τ ∗ = O NB

(1.78)

and

respectively. If we write   ν1 ν2    ν =  ..  = O τ . νt then Eq. (1.77) and (1.78) reduce to δi νi = ai r− k and δi νi∗ = bi respectively, where a = (a1 , a2 , . . . , at ) = O Q

36


and b = (b1 , b2 , . . . , bt ) = O N B We see then that we have both intrablock and interblock estimators of νi if δi is not equal to 0 or to rk. For the component νt , only the interblock estimator exists. If other roots are equal to rk, then the intrablock estimators for the corresponding treatment parameters do not exist. Similarly, if other roots are equal to zero, then the interblock estimators for the corresponding treatment parameters do not exist. The treatment parameters ν1 , ν2 , . . . , νt−1 are necessarily treatment contrasts. If the design is connected, then no δi (i = 1, 2, . . . , t − 1) will equal rk. The combined NE (1.76) are now transformed to ' ( δi bi 1 r− ν i = ai + + δi (i = 1, 2, . . . , t) (1.79) k k(1 + γ k) k(1 + γ k) We know that 1 var(Q) = rI − N N σe2 k

and var(N B) = N var(B)N k = N 2 (1 + γ k)I N σe Hence

δi var(a) = σe2 diag r − k

and var(b) = σe2 k(1 + γ k)diag(δi ) So we see from (1.79) that combined estimation of the parameter vector ν consists of combining intrablock and interblock estimators of components of ν, weighting inversely as their variances.

1.10 ESTIMATION OF WEIGHTS FOR THE COMBINED ANALYSIS The estimator for the treatment effects as given by (1.61) depends on the weights w and wj as can be seen from (1.58) and (1.59). If these weights were known,

37

ESTIMATION OF WEIGHTS FOR THE COMBINED ANALYSIS

or alternatively as is apparent from (1.68), if the ratios of the interblock variance and the intrablock variance, ρj =

σe2 + kj σβ2 w = wj σe2

were known, then the solution (1.63) to the NE (1.55) would lead to best linear unbiased estimators for estimable functions of treatment effects. Usually, however, these parameters are not known and have to be estimated from the data. If the estimators are used instead of the unknown parameters, then the solutions to the normal equations (1.55) lose some of their good properties. It is for this reason that the properties of the combined estimator have to be examined critically, in particular with regard to their dependence on the type of estimator for the ρj ’s, and with regard to the question of how the combined estimator compares with the intrablock estimator. Before we discuss these questions in some more detail, we shall outline the “classical” procedure for estimating the ρj . Since this method was proposed first by Yates (1940a) we shall refer to it as the Yates procedure or to the estimators as the Yates estimators. 1.10.1 Yates Procedure One way to estimate w and wj is to first estimate σe2 and σβ2 and then use these estimators to estimate w and wj . If the estimators are denoted by σe2 and σβ2 , respectively, then we estimate w and wj as w =

1 σe2

w j =

σe2

1 + kj σβ2

(j = 1, 2, . . . , b)

(1.80)

Obviously, from Table 1.1 σe2 = MS(I |I, Xβ , Xτ )

(1.81)

To estimate σβ2 we turn to Table 1.2. Under model (1.50) with covariance structure (1.51) we find [see also (1.39)] 

 1 E[SS(X β |I, X τ )] = (b − 1)σe2 + n − n2  σβ2 ri ij

(1.82)

ij

Hence it follows from (1.81) and (1.82) that σβ2 =

b−1 1 2 n− ri nij

! SS(Xβ |I, Xτ ) − SS(I |I, Xβ , Xτ )

(1.83)

38


The estimators (1.81) and (1.83) are then substituted into (1.80) to obtain w and w j (j = 1, 2, . . . , b). If in (1.83) σβ2 ≤ 0 for a given data set, we take = w j = w

1 MS(I |I, Xβ , Xτ )

In either case w and w j are substituted into (1.58) and (1.59) and hence into τ i is estimated by substituting w the solution (1.61). Also, var τi − , w j and 2 σe into (1.62) and (1.64). For alternative estimation procedures see Section 1.11, and for a numerical example see Section 1.14.3. 1.10.2 Properties of Combined Estimators As we have already pointed out, the fact that the unknown parameters in (1.61) are replaced by their estimators will have an effect on the properties of the estimators for treatment effects. The two properties we are concerned about here are unbiasedness and minimum variance. Let c τ be an estimable function of the treatment effects, let t = c τ be its intrablock estimator, t (ρ) = c τ its combined (Aitken) estimator with ρ known (for the present discussion we shall confine ourselves to proper designs), and t ( ρ ) = c τ the combined estimator when in (1.68) ρ is replaced by ρ = w / w . Roy and Shah (1962) have shown that for the general incomplete block design, although the Yates procedure leads to a biased estimator for ρ, the estimators for treatment contrasts obtained by the method just described are unbiased, that is, E [t ( ρ )] = c τ With regard to var[t ( ρ )], it is clear that due to sampling fluctuations of ρ we have var[t (ρ)] < var[t ( ρ )] that is, the combined estimators no longer have minimum variance. The crucial question in this context, however, is: When is var[t ( ρ )] < var(t)? In other words: When is the combined estimator more efficient than the intrablock estimator? The answer to this question depends on several things such as (1) the true value of ρ, (2) the type of estimator for ρ, and (3) the number of treatments and blocks. It is therefore not surprising that so far no complete answer has been given. The most general result for the Yates estimator (and a somewhat larger class of estimators) is that of Shah (1964) based upon some general results by Roy and Shah (1962) [see also Bhattacharya (1998) for proper designs]. It is shown there that the combined estimator for any treatment contrast in any (proper) incomplete block design has variance smaller than that of the corresponding intrablock estimator if ρ does not exceed 2, or, equivalently, if σβ2 ≤ σe2 /k. This

39

MAXIMUM-LIKELIHOOD TYPE ESTIMATION

is a condition that, if blocking is effective at all, one would not in general expect to be satisfied. The problem therefore remains to find methods of constructing estimators for ρ such that the combined estimator for treatment contrasts is uniformly better than the corresponding intrablock estimator, in the sense of having smaller variances for all values of ρ. For certain incomplete block designs this goal has been achieved. We shall mention these results in the following chapters. The only general advice that we give at this point in conjunction with the use of the Yates estimator is the somewhat imprecise advice to use the intrablock rather than the combined estimator if the number of treatments is “small.” The reason for this is that in such a situation the degrees of freedom for MS(I |I, X β , Xτ ) and MS(Xβ |I, Xτ ) are likely to be small also, which would imply that σe2 and σβ2 , and hence ρ, cannot be estimated very precisely.

1.11 MAXIMUM-LIKELIHOOD TYPE ESTIMATION In this section we discuss alternatives to the Yates procedure (see Section 1.10) of estimating the variance components σβ2 and σe2 for the combined analysis. These estimators are maximum-likelihood type estimators. This necessitates the assumption of normality, which is not in agreement with our underlying philosophy of finite population randomization analysis. The reason for discussing them, however, is the fact that they can easily be implemented in existing software, in particular SAS PROC MIXED (SAS, 1999–2000) (see Section 1.14). 1.11.1 Maximum-Likelihood Estimation It is convenient to rewrite model (1.49) with its covariance structure (1.51) in matrix notation as follows: y = µI + X τ τ + X β β + e = Xα + U β + e

(1.84)

where Xα represents the fixed part, with X = (I Xτ ), α = (µ, τ ), and U β + e represents the random part. Thus

and

E(y) = Xα var(y) = U U σβ2 + I n σe2 = V σe2

(1.85)

40


with V = γ U U + I n and γ = σβ2 /σe2 [see also (1.53), (1.54)]. We then assume that y ∼ Nn (Xα, V σe2 )

(1.86)

that is, y follows as a multivariate normal distribution (see I.4.17.1). The logarithm of the likelihood function for y of (1.86) is then given by λ = − 12 n log π − 12 nσe2 −

1 2

log|V | − 12 (y − Xα) V −1 (y − Xα)/σe2

(1.87)

Hartley and Rao (1967) show that the maximum-likelihood (ML) estimator of α, γ , and σe2 are obtained by solving the following equations: 1 (X V −1 y − X V −1 Xα) = 0 (1.88) σe2 1 1 − tr(V −1 U U ) + 2 (y − Xα) V −1 U U V −1 (By − Xα) = 0 2 2σe −

1 n + 4 (y − Xα) V −1 (y − Xα) = 0 2 2σe 2σe

where tr(A) represents the trace of the matrix A. The basic feature of this method is that the fixed effects and the variance components associated with the random effects are estimated simultaneously in an iterative procedure. We shall not go into the details of the numerical implementation (see, e.g., Hemmerle and Hartley, 1973), but refer to the example given in Section 1.14.4 using SAS. 1.11.2 Restricted Maximum-Likelihood Estimation Specifically for the estimation of weights for the purpose of recovery of interblock information, Patterson and Thompson (1971) introduced a modified maximumlikelihood procedure. The basic idea is to obtain estimators for the variance components that are free of the fixed effects in the sense that the likelihood does not contain the fixed effect. Operationally this is accomplished by dividing the likelihood function (1.87) into two parts, one being based on treatment contrasts and the other being based on error contrasts, that is, contrasts with expected value zero. Maximizing this second part will lead to estimates of functions of σβ2 and σe2 . Because of the procedure employed, these estimates are by some referred to as residual maximum-likelihood estimates (REML), by others as restricted maximum likelihood estimates (REML). The latter name is derived from the fact that maximizing the part of the likelihood free of the fixed effects can be thought of as maximizing the likelihood over a restricted parameter set, an idea first proposed by Thompson (1962) for random effects models and generalized for the general linear mixed model by Corbeil and Searle (1976), based on the

41

MAXIMUM-LIKELIHOOD TYPE ESTIMATION

work by Patterson and Thompson (1971). We shall give below a brief outline of the basic idea of REML estimation following Corbeil and Searle (1976). Consider model (1.84) and assume that the observations are ordered by treatments, where the ith treatment is replicated ri times (i = 1, 2, . . . , t). Then the matrix X can simply be written as  Ir1   X=  0

 Ir2 ..

.

t 0  + Ir i ≡  i=1 Irt

(1.89)

To separate the log-likelihood function (1.87) into the two parts mentioned above, we employ the transformation (as proposed by Patterson and Thompson, 1971) . y [S ..V −1 X]

(1.90)

where S = I − X(X X)−1 X =

t +

I ri −

i=1

1 Ir I ri i ri

(1.91)

is symmetric and idempotent. Furthermore, SX = 0, and hence Sy is distributed N (0, SV Sσe2 ) independently of X V −1 y. It follows from (1.91) that S is singular. Hence, instead of S we shall use in (1.90) a matrix, T say, which is derived from S by deleting its r1 th, (r1 + r2 )th, (r1 + r2 + r3 )th, . . ., (r1 + r2 + · · · + rt )th rows, thereby reducing an n × n matrix to an (n − t) × n matrix (with n − t representing the number of linearly independent error contrasts). More explicitly, we can write T as T =

t ' + i=1

=

t ' + i=1

. 1 I ri −1 .. 0ri −1 − Iri −1 Iri ri I ri −1 −

(

. 1 1 Iri −1 Iri −1 .. − Iri −1 ri ri

( (1.92)

It follows from (1.89) and (1.92) that TX = 0

(1.93)

42


Considering now the transformation z=

T X V −1

y

it follows from (1.86) and (1.93) that 0 T V T σe2 , z∼N X V −1 Xα 0

0

X V −1 Xσe2

(1.94)

Clearly, the likelihood function of z consists of two parts, one for T y, which is free of fixed effects, and one for X V −1 y pertaining to the fixed effects. In particular, the log likelihood of T y then is λ1 = 12 (n − t)log 2π − 12 (n − t)log σe2 − 12 log |T V T | − 12 y T (T V T )−1 y/σe2

(1.95)

The REML estimators for γ = σβ2 /σe2 and σe2 are obtained by solving the equations ∂λ1 =0 ∂γ

(1.96)

∂λ1 =0 ∂σe2

(1.97)

The resulting equations have no analytic solutions and have to be solved iteratively. We denote the solutions, that is, estimates by γ and σe2 , respectively. The fixed effects, represented by α in (1.84), can be estimated by considering the log likelihood of X V −1 y, which is given by λ2 = − 12 tlog 2π − 12 t log σe2 − 12 log | X V −1 X| − 12 (y − Xα) V −1 X(X V −1 X)−1 X V −1 (y − Xα)/σe2

(1.98)

Solving ∂λ2 =0 ∂α leads to the estimator α = (X V −1 X)−1 X V −1 y

(1.99)

43

EFFICIENCY FACTOR OF AN INCOMPLETE BLOCK DESIGN

This, of course, assumes that V is known. Since it is not, we substitute γ from (1.96) and (1.97) for γ in (1.99) and obtain the estimate α = (X V

−1

X)−1 X V

−1

y

(1.100)

denotes V with γ replaced by where V γ. An approximate estimate of the variance of α is given by v) ar( α) ∼ = (X V

−1

X)−1 σe2

(1.101)

For a numerical example using REML see Section 1.14.4.

1.12 EFFICIENCY FACTOR OF AN INCOMPLETE BLOCK DESIGN We have seen in Sections I.9.3 and I.10.2.9, for example, how we can compare different error control designs with each other by using the notion of relative efficiency. In this case, we compare two error control designs after we have performed the experiment using a particular error control design. For example, after we have used an RCBD we might ask: How would we have done with a corresponding CRD? In other cases, however, we may want to compare error control designs before we begin an experiment. In particular, we may want to compare an incomplete block design (IBD) with either a CRD or an RCBD, or we may want to compare competing IBDs with each other. For this purpose we shall use a quantity that is referred to as the efficiency factor of the IBD. It compares, apart from the residual variance, σe2 , the average variance of simple treatment comparisons for the two competing designs.

1.12.1 Average Variance for Treatment Comparisons for an IBD Let us now consider av. var( τi − τi )

i=i

(1.102)

for a connected IBD. We know, of course, that (1.102) is a function of C − , a generalized inverse of the C matrix. Suppose now that all the block sizes are equal to k. Then we have 1 C = R − NN k

44


and we know that C√has one zero root, dt = 0 say, with associated normalized eigenvector ξ t = (1/ t)I. Let the other roots be d1 , d2 , . . . , dt−1 with associated orthonormal eigenvectors ξ 1 , ξ 2 , . . . , ξ t−1 . Then ξ i C = di ξ i

(i = 1, 2, . . . , t − 1)

and from ξ i Cτ = di ξ i τ it follows that 1 ξ Q ξ* iτ = di i and τ) = var(ξ i

1 1 ξ i Cξ i σe2 = σe2 2 d di i

(1.103)

√ Using the fact that ξ t = (1/ t)I, that ξ 1 , ξ 2 , . . . , ξ t−1 are mutually perpendicular and perpendicular to ξ 1 , and that t

ξ i ξ i = I

i=1

we have with z = (z1 , z2 , . . . , zt ) t−1

(ξ i z)2

=z

t−1

i=1

ξ i ξ i

z

i=1

= z (I − ξ t ξ t )z

=

t i=1

=

t

t 2 1 zi2 − zi t i=1

(zi − z)2

(1.104)

i=1

It is also easy to verify that t 2 1 (zi − zi )2 = (zi − z)2 t (t − 1) t − 1 i=i

i=1

(1.105)

45


Taking zi = τi − τi , substituting into (1.105) using (1.104) and then taking expectations and using (1.103), yields for (1.102) t−1 2 1 2 av. var( τi − τ )= σ t −1 di e i=i i

(1.106)

i=1

1.12.2 Definition of Efficiency Factor It is natural in attempting to evaluate the efficiency of an IBD to compare it with a CRD since this is always a possible competing design. For a CRD with ri replications for treatment i, the average variance of treatment differences is 1 2 2 1 2 av. var( τi − τi ) = av. + σe(CRD) = σ (1.107) ri ri r h e(CRD) i=i i=i where r h is the harmonic mean of the ri , that is, 1 1 1 = rh t ri i

We shall digress here for a moment and show that the best CRD is the one with all ri = r, and that is the design with which we shall compare the IBD. For this and later derivations we need the “old” result that the harmonic mean of a set of positive numbers is not greater than the arithmetic mean. It seems useful to give an elementary proof of this. Let the set of numbers be {xi , i = 1, 2, . . . , m}. Consider the quadratic q(β) =

m √ i=1

1 xi − β √ xi

2

Clearly q(β) ≥ 0 for all β. The minimizing value of β is obtained by using least squares which gives the NE 1 = m β xi i

The minimum sum of squares is

xi − βm

i

Hence i

m2 xi − ≥0 1 i

xi

46

or


1 xi m i

1 1 m xi

≥1

i

or x ≥1 xh with equality if and only if xi = x for all i. This result implies that the best CRD will have ri = r and r = n/t where n is the total numbers of EUs. This can happen, of course, only if n/t is an integer. If n/t is not an integer so that n = pt + q (0 < q < t), then the best CRD will have q treatments replicated p + 1 times. Consider now the case of an IBD with b blocks of size k and ri replications for the ith treatment. Then the total number of EUs is n = bk = ri . Suppose also that n = rt, so that an equireplicate CRD is possible. The average variance 2 for such a design is 2σe(CRD) /r, whereas the average variance for the IBD is 2 2σe(IBD) /c where, as shown in (1.106), c is the harmonic mean of the positive eigenvalues of R − (1/k) NN (see Kempthorne, 1956). It is natural to write 2 2 = σe(IBD) we have c = rE, so that with σe(CRD) 2/r av. var( τi − τi )CRD = =E av. var( τi − τi )IBD 2/rE

(1.108)

The quantity E thus defined is called the efficiency factor of the IBD. It is clearly a numerical property of the treatment-block configuration only and hence a characteristic of a given IBD. We add the following remarks: 1. The same definition of E in (1.108) could have been obtained by using the average variance for an RCBD with b = r blocks instead of the average 2 2 = σe(IBD) . variance for an equireplicate CRD assuming that σe(RCBD) 2. Although E is a useful quantity to compare designs, it does not, of course, give the full story. It compares average variances only under the assump2 tion of equality of residual variances, whereas we typically expect σe(IBD) < 2 2 2 σe(CRD) and σe(IBD) < σe(RCBD) . 3. The efficiency factor pertains only to the intrablock analysis and ignores the interblock information. 4. Each IBD will have associated with it an efficiency factor E. In order to compare two competing IBDs with the same n and with efficiency factors E1 and E2 , respectively, we would typically choose the one with the higher E value.

47


1.12.3 Upper Bound for the Efficiency Factor Using again the fact that the harmonic mean of positive numbers is not greater than the arithmetic mean, we have

(t − 1)c ≤

t−1 i=1

1 di = trace R − N N k =

t

ri −

ij

i=1

=n−

1 2 nij k

1 2 nij k

(1.109)

ij

The largest value of the right-hand side of (1.109) is obtained for the smallest value of ij n2ij . Since nij is one of the numbers 0, 1, 2, . . . , k, the minimum 2 value of nij will be achieved when n of the nij ’s are 1 and the remaining are zero. Since then n2ij = nij and j nij = ri , it follows from (1.109) that (t − 1)c ≤

ri −

i

1 k−1 tr ri = k k i

or, since c = rE, E≤

(k − 1)t/(t − 1)k r/r

But since tr = n = tr, we have finally E≤

(k − 1)t (t − 1)k

(1.110)

Since for an IBD k < t, we can write further E≤

(k − 1)t |t|

11.2750000 16.9000000 23.4000000 26.5250000

1.9774510 2.6632921 2.6632921 1.9774510

0.0294 0.0239 0.0127 0.0055

Dependent Variable: Y Parameter TRT1 - (TRT2+TRT3)/2 TRT1 - TRT4 TRT2 - TRT3

Estimate

Standard Error

t Value

Pr > |t|

-8.8750000 -15.2500000 -6.5000000

2.61157280 3.01558452 4.26468053

-3.40 -5.06 -1.52

0.0768 0.0369 0.2670

58


5. The top part of the ANOVA table provides the partition SS(MODEL) + SS(ERROR) = SS(TOTAL) and from that produces MS(ERROR) = σe2 = 9.094 6. The lower part of the ANOVA table provides type I SS (sequential SS for ordered model; see I.4.7.2) and type III SS (partial SS). From the latter we obtain the P value (.0975) for the test of H0 : τ1 = τ2 = · · · = τt versus H1 :

not all τi are the same

We stress that the P value for blocks should be ignored (see I.9.2). 7. The e-option of LSMEANS gives the coefficients for the solution vector to compute the treatment least squares means, for example, βi∗ + τ1∗ LSMEAN(TRT 1) = µ∗ + .2 = 31.125 + .2(−7.6875 − 4.0625 − 1.9375 − 9.3125 + 0) − 15.25 = 11.275 The standard error is computed by making use of the G inverse (see item 2) and MS(ERROR). 8. The t tests are performed for the prespecified contrasts among the leastsquares means; for example, TRT2-TRT3 = τ2∗ − τ3∗ = 9.625 + 3.125 = − 6.5 se(τ2∗ − τ3∗ ) = [(1.25 + 1.25 − 2 × .25) × 9.094]1/2 = 4.265 t=

− 6.5 = −1.52 4.265

1.14.2 Intrablock Analysis Using the Absorb Option in SAS PROC GLM A computational method as described in Section 1.3 using the RNE can be implemented in SAS PROC GLM by using the ABSORB OPTION. This is illustrated in Table 1.9.

59

COMPUTATIONAL PROCEDURES

Table 1.9 Intra Block Analysis Using Reduced Normal Equations With Post-hoc Comparisons proc glm data=IBD; class BLOCK TRT; absorb BLOCK; model Y = TRT/XPX inverse solution e; estimate 'TRT1 - (TRT2+TRT3)/2' TRT 1 -.5 -.5 0; estimate 'TRT1 - TRT4' TRT 1 0 0 -1; estimate 'TRT2 - TRT3' TRT 0 1 -1 0; title1 'TABLE 1.9'; title2 'INTRA-BLOCK ANALYSIS USING REDUCED NORMAL EQUATIONS'; title3 'WITH POST-HOC COMPARISONS'; run; The GLM Procedure Class Level Information Class

Levels

Values

BLOCK

5

1 2 3 4 5

TRT

4

1 2 3 4

Number of observations 10 The X'X Matrix

TRT TRT TRT TRT Y

1 2 3 4

TRT 1

TRT 2

TRT 3

TRT 4

Y

1.5 -0.5 -0.5 -0.5 -16.5

-0.5 1 0 -0.5 -2

-0.5 0 1 -0.5 4.5

-0.5 -0.5 -0.5 1.5 14

-16.5 -2 4.5 14 275

The GLM Procedure X'X Generalized Inverse (g2)

TRT TRT TRT TRT Y

1 2 3 4

TRT 1

TRT 2

TRT 3

TRT 4

Y

1 0.5 0.5 0 -15.25

0.5 1.25 0.25 0 -9.625

0.5 0.25 1.25 0 -3.125

0 0 0 0 0

-15.25 -9.625 -3.125 0 18.1875

General Form of Estimable Functions Given that the coefficients for all absorbed effects are zero

60


Table 1.9 (Continued ) Effect

Coefficients

TRT TRT TRT TRT

1 2 3 4

L1 L2 L3 -L1-L2-L3

Dependent Variable: Y Source

DF

Sum of Squares

Mean Square

F Value

Pr > F

8.14

0.1137

Model

7

518.2125000

74.0303571

Error

2

18.1875000

9.0937500

Corrected Total

9

536.4000000

R-Square

Coeff Var

Root MSE

Y Mean

0.966093

15.54425

3.015585

19.40000

Source

DF

BLOCK TRT Source TRT

Type I SS

Mean Square

F Value

Pr > F

4 3

261.4000000 256.8125000

65.3500000 85.6041667

7.19 9.41

0.1259 0.0975

DF

Type III SS

Mean Square

F Value

Pr > F

3

256.8125000

85.6041667

9.41

0.0975

Parameter TRT1 - (TRT2+TRT3)/2 TRT1 - TRT4 TRT2 - TRT3

Parameter TRT TRT TRT TRT

Estimate

Standard Error

t Value

Pr > |t|

-8.8750000 -15.2500000 -6.5000000

2.61157280 3.01558452 4.26468053

-3.40 -5.06 -1.52

0.0768 0.0369 0.2670

Estimate 1 2 3 4

-15.25000000 -9.62500000 -3.12500000 0.00000000

B B B B

Standard Error

t Value

Pr > |t|

3.01558452 3.37152599 3.37152599 .

-5.06 -2.85 -0.93 .

0.0369 0.1039 0.4518 .

NOTE: The X'X matrix has been found to be singular, and a generalized inverse was used to solve the normal equations. Terms whose estimates are followed by the letter 'B' are not uniquely estimable.

61


We make the following comments about the SAS output: 1. The X X matrix is now the C-matrix of (1.9). 2. The X X generalized inverse is obtained by the SAS convention of setting −1 of (1.16). τ4∗ = 0. This g inverse is therefore different from C 3. The ANOVA table provides the same information as in Table 1.8, except that it does not give solutions for the intercept and blocks. Hence, this analysis cannot be used to obtain treatment least-squares means and their standard error. 1.14.3 Combined Intra- and Interblock Analysis Using the Yates Procedure Using SAS PROC VARCOMP illustrates the estimation of σβ2 according to the method described in Section 1.10.1. The result is presented in Table 1.10. The option type I produces Table 1.2 with % & E[MS(BLOCK)] = E MS(X β |I, X τ ) as given in (1.82). This yields σe2 = 9.09 (as in Table 1.8) and σβ2 = 7.13. 9.09 + 2 × 7.13 Substituting ρ= = 2.57 into (1.60) we obtain 9.09  2.0814 −0.3062 −0.3062 −0.3062  1.3877 0.0000 −0.3062  = −0.3062 A −0.3062 0.0000 1.3877 −0.3062 −0.3062 −0.3062 −0.3062 2.0814 

and 

−1 A

0.541654 0.146619 = 0.146619 0.122823

0.146619 0.785320 0.064704 0.146619

0.146619 0.064704 0.785320 0.146619

and  4.6242   =  8.8528 P 22.1358 39.5816 

 0.122823 0.146619  0.146619 0.541654

62 DF 3 4 2 9

Source

TRT BLOCK Error Corrected Total

5 4

BLOCK TRT

1 2 3 4

1 2 3 4 5

Values

Var(BLOCK) Var(Error)

Var(Error) + 0.6667 Var(BLOCK) + Q(TRT) Var(Error) + 1.5 Var(BLOCK) Var(Error) .

Expected Mean Square

7.12847 9.09375

Estimate

Type 1 Estimates

146.355556 19.786458 9.093750 .

Mean Square

Variance Component

439.066667 79.145833 18.187500 536.400000

Sum of Squares

Type 1 Analysis of Variance

Dependent Variable: Y

Number of observations 10

Levels

Class

Class Level Information

title1 'TABLE 1.10'; title2 'ESTIMATION OF BLOCK VARIANCE COMPONENT'; title3 'USING THE YATES PROCEDURE'; run; Variance Components Estimation Procedure

class BLOCK TRT; model Y = TRT BLOCK/fixed=1;

proc varcomp data=IBD method=type1;

Table 1.10 Estimation of Block Variance Component Using the Yates Procedure

63


We then obtain

 11.9097   −1 P = 14.8659 τ =A 24.4379 26.5510 

We note that the elements of τ are actually the treatment least-squares means. Their estimated variances and the estimated variances for the treatment contrasts −1 × 9.09 (see Tables 1.13 and 1.14). are obtained from A 1.14.4 Combined Intra- and Interblock Analysis Using SAS PROC MIXED We illustrate here the numerical implementation of the ML and REML procedures as described in Sections 1.11.1 and 1.11.2, respectively, using SAS PROC MIXED. The results of the ML estimation are given in Table 1.11. σe2 = It takes four interations to obtain the solutions, yielding σβ2 = 7.45 and 4.14 and hence γ = 1.80 (Notice that these are quite different from the estimates obtained by the Yates procedure (Section 1.14.3) and the REML procedure as given below). Since SAS uses a different parametrization than the one used in (1.84) it obtains “estimates” of µ and τi (i = 1, 2, 3, 4) separately. The type 3 coefficients indicate that the solutions for µ, τi (i = 1, 2, 3, 4) are actually estimates of µ + τ4 , τ1 − τ4 , τ2 − τ4 , τ3 − τ4 , respectively. From these solutions the least-squares means are then obtained as LSMEAN(TRT 1) = µ + τ1

=

11.65

=

α1

LSMEAN(TRT 2) = µ + τ2

=

15.63

=

α2

LSMEAN(TRT 3) = µ + τ3

=

24.09

=

α3

LSMEAN(TRT 4) = µ

=

26.53

=

α4

where the αi denote the solutions to (1.87). The REML procedure is illustrated in Table 1.12. It takes three iterations to σe2 = 10.17, and hence γ = 0.62. We note obtain the estimates σβ2 = 6.35 and that the REML and ML least-squares means are numerically quite similar even though γ and σe2 are substantially different from γ and σe2 , respectively. 1.14.5 Comparison of Estimation Procedures For a small proper incomplete block design we have employed the above four methods of estimating treatment least-squares means and treatment comparisons: M1: Intrablock analysis M2: Combined analysis: Yates

64


Table 1.11 Combined Analysis Using Maximum Likelihood With Post-hoc Comparisons proc mixed data=IBD method=ML; class BLOCK TRT; model Y = TRT/ solution E3 ddfm=Satterth; random BLOCK; lsmeans TRT; estimate 'TRT1 - (TRT2+TRT3)/2' TRT 1 -.5 -.5 0; estimate 'TRT1 - TRT4' TRT 1 0 0 -1; estimate 'TRT2 - TRT3' TRT 0 1 -1 0; title1 'TABLE 1.11'; title2 'COMBINED ANALYSIS USING MAXIMUM LIKELIHOOD'; title3 'WITH POST-HOC COMPARISONS'; run; The Mixed Procedure Model Information Data Set Dependent Variable Covariance Structure Estimation Method Residual Variance Method Fixed Effects SE Method Degrees of Freedom Method

WORK.IBD Y Variance Components ML Profile Model-Based Satterthwaite

Class Level Information Class

Levels

BLOCK TRT

5 4

Values 1 2 3 4 5 1 2 3 4

Dimensions Covariance Parameters Columns in X Columns in Z Subjects Max Obs Per Subject Observations Used Observations Not Used Total Observations

2 5 5 1 10 10 0 10

65


Table 1.11 (Continued ) Iteration History Iteration

Evaluations

-2 Log Like

Criterion

0 1 2 3 4

1 2 1 1 1

51.13433487 50.29813032 50.22541829 50.22033455 50.22029773

0.00396161 0.00030283 0.00000230 0.00000000

Convergence criteria met. Covariance Parameter Estimates Cov Parm

Estimate

BLOCK Residual

7.4528 4.1426 Fit Statistics

-2 Log Likelihood AIC (smaller is better) AICC (smaller is better) BIC (smaller is better)

50.2 62.2 90.2 59.9

Solution for Fixed Effects

Effect

TRT

Estimate

Standard Error

DF

t Value

Pr > |t|

Intercept TRT TRT TRT TRT

1 2 3 4

26.5329 -14.8822 -10.9029 -2.4380 0

1.7767 1.9330 2.1495 2.1495 .

8.87 4.75 4.82 4.82 .

14.93 -7.70 -5.07 -1.13 .

F

3

4.76

23.37

0.0028

Estimates

Label

Estimate

Standard Error

DF

t Value

Pr > |t|

TRT1 - (TRT2+TRT3)/2 TRT1 - TRT4 TRT2 - TRT3

-8.2117 -14.8822 -8.4650

1.7085 1.9330 2.6087

4.3 4.75 5.86

-4.81 -7.70 -3.24

0.0072 0.0007 0.0182

Least Squares Means

Effect

TRT

TRT TRT TRT TRT

1 2 3 4

Estimate

Standard Error

DF

t Value

Pr > |t|

11.6506 15.6299 24.0949 26.5329

1.7767 2.0786 2.0786 1.7767

8.87 9.98 9.98 8.87

6.56 7.52 11.59 14.93

0.0001 |t|

11.9914 14.6444 24.5291 26.5596

2.2615 2.7365 2.7365 2.2615

5.93 5.52 5.52 5.93

5.30 5.35 8.96 11.74

0.0019 0.0023 0.0002 λ

(2.5)

b≥t

(2.6)

Relationship (2.3) follows immediately from the fact that the total number of observation for the BIB design is    nij  = nij n= i

j

and to this we apply (2.1) and (2.2).

j

i

73

PROPERTIES OF BIB DESIGNS

To establish (2.4) we consider a fixed treatment that is replicated r times and in each of these r blocks there are k − 1 other treatments. On the other hand, because of (iii) in Definition 2.1, each of the remaining t − 1 treatments must occur in these blocks exactly λ times. This leads immediately to (2.4). Condition (2.5) follows from (2.4) and the fact that for an incomplete block design k < t. Fisher (1940) first proved (2.6) to be a necessary condition for the existence of a balanced incomplete block design. The following proof is based on a property of the matrix N N . From the definition of the BIB design it follows that 

r

·

 λ ·   ·   = (r − λ)I + λII λ   ·

λ

r

λ

λ   ·  NN =  ·   · λ

r ·

· ·

·

(2.7)

Then N N I = (r − λ)I + tλI = (r − λ + λt)I √ So (1/ t)I is a normalized eigenvector with root √ r − λ + λt. If ξ is another eigenvector of NN , which is orthogonal to (1/ t)I, then we have from (2.7) that NN ξ = (r − λ)ξ and hence (r − λ) is an eigenvalue with multiplicity t − 1. It follows then that detNN = (r − λ)t−1 (r − λ + λt) or, using (2.4), detN N = (r − λ)t−1 rk

(2.8)

Because of (2.5), this determinant is positive definite of rank t. Now t = rank N N = rank N ≤ min(t, b). Hence b ≥ t. In connection with the above necessary conditions for the existence of a BIB design, we shall mention briefly the special case when t = b and hence r = k. Definition 2.2 An incomplete block design (IBD) with the same number of treatments and blocks is called a symmetrical incomplete block design.

74

BALANCED INCOMPLETE BLOCK DESIGNS

For a symmetrical BIB design, (2.8) will be of the form detN N = (detN )2 = (r − λ)t−1 r 2 This implies, for example, that if t is even then r − λ must be a perfect square for a symmetrical BIB design to exist. This is only one of several necessary conditions (Shrikhande, 1950) for the existence of a symmetrical BIB design. Nonexistence theorems and methods of actually constructing BIB designs for a given set of parameters have become an important subject in itself, and we shall treat some aspects of this in Chapter 3. Before we consider the analysis of BIB designs, we give as an example the design mentioned in Section 1.1. Example 2.1 We have t = 7, b = 7, r = 4, k = 4, λ = 2 and denoting the treatments now by 0, 1, . . . , 6 the design is listed as (0, 3, 6, 5) (2, 5, 4, 6) (6, 0, 1, 4) (0, 1, 2, 5) (1, 6, 2, 3) (4, 2, 3, 0) (1, 3, 4, 5) where each row represents a block, the treatments in each block being randomly assigned to the experimental units. 2.4 ANALYSIS OF BIB DESIGNS 2.4.1 Intrablock Analysis Using the results of Section 1.3.2 the RNE (1.7) take on the form   k−1 λ λ r − −  k     k k    τ1 Q1   ..      λ k−1  τ  Q   − r .     2  2  k   =   .k .. λ   ..   ..   .  . . −  .   .   k    τt Qt  λ λ k − 1 − − r k k k

75

ANALYSIS OF BIB DESIGNS

or r

k−1 λ λ + τ =Q I − II k k k

(2.9)

A solution to (2.9) is obtained by using Theorem 1.1 with a = λ/k. Then −1 = C

1 I k−1 λ r + k k

(2.10)

Using (2.4) in (2.10) we can write r k−1 λ k−1 (k − 1)t + = r k−1+ =r k k k t −1 k(t − 1) and hence −1 = C

1 I (k − 1)t r k(t − 1)

(2.11)

−1 σ 2 is the variance–covariance matrix for estimable functions of Recall that C e treatment effects. Thus, from (2.11) we have var ( τi − τi ) =

2 σ2 (k − 1)t e r k(t − 1)

for all i = i . Therefore also τi − τi ) = av. var(

i=i

2 σ2 (k − 1)t e r k(t − 1)

It follows then from (1.89) that (k − 1)t/k(t − 1) is the efficiency factor E of the BIB design and according to (1.91) the upper bound for any IBD. We then write −1 = 1 I C rE

(2.12)

Hence τ=

Q rE

(2.13)

76


is a solution to (2.9). Furthermore, it followsfrom (2.13) and (1.20) that any linear estimable function c τ = i ci τi with ci = 0 is estimated by

ci

Qi rE

(2.14)

ci2

σe2 rE

(2.15)

i

with variance, using (2.12), i

The intrablock error variance σe2 is estimated in the usual way by using the analysis of variance given in Table 1.2. 2.4.2 Combined Analysis It follows from the definition of the BIB design and the form of NN as given in (2.7) that the coefficient matrix, A say, in (1.60) is now of the form r λ λ −1 A = r − (1 − ρ ) (2.16) − I − (1 − ρ −1 ) II k k k By a familiar result in matrix algebra (e.g., Graybill, 1969) we find   λ −1 (1 − ρ )   1 k  II  A−1 = I +   r r λ λ r − (1 − ρ −1 ) r − (1 − ρ −1 ) − + (t − 1) k k k k (2.17)

Denoting the (i, i ) element of A−1 by a ii , it follows from (1.62) and (1.63) that the linear function of treatment effects, c τ , with ci = 0 is estimated by c τ with variance var ci c2 a ii σe2 + ci ci a ii σe2 (2.18) τi = i

i

i

i,i i=i

Substituting the elements of (2.17) into (2.18) yields σe2 var c i ci2 τi = r λ −1 i i r − 1−ρ − k k

(2.19)

77

ESTIMATION OF ρ

or, substituting (1.65) in (2.19), var

ci τi

=

i

i

ci2

σe2 + kσβ2 σe2

rσe2 + (rk − r + λ)σβ2

(2.20)

The following special familiar cases can be deduced easily from (2.19): 1. r = λ (i.e., RCBD): var

τi ci

=

i

ci2

σe2 r

ci2

σe2 rE

i

2. ρ −1 = 0: var

i

ci τi

=

i

which is the same as (2.15), implying that no interblock information exists. 3. ρ −1 = 1: σ2 var ci ci2 e τi = r i

i

which is the same as the variance for a CRD. We note that σe2 here refers to the error variance in a CRD, whereas σe2 in case 1 refers to the within-block variance.

2.5 ESTIMATION OF ρ As we already know from the discussion in the previous chapter and as is obvious from τ for c τ depends on w = 1/σe2 and w = combined estimator c (2.17), the 1/ σe2 + kσβ2 through ρ = w/w . Since, in general, these parameters are not known, they will have to be estimated from the data; the estimates will have to be substituted into (2.17) in order to obtain the estimator for c τ , which we shall denote by c τ . Note that in our notation ci τ depends on τ depends on ρ and c ρ , where ρ is an estimator for ρ. In the previous chapter we described one method of estimating ρ, the Yates procedure, which estimates σe2 and σβ2 from the intrablock analyses of variance with the model assuming the block effects to be random variables. This is a general procedure and as such applicable also to BIB designs. However, we also mentioned the problem that arises when using the Yates estimators, since one is

78


not assured that the combined estimator for c τ is better than the corresponding intrablock estimator. The following theorem, which is a special case of a more general theorem given by Shah (1964), deals with this problem and gives a solution under a nonrestrictive condition. Let t (t − 1)k(k − 1) 1 Z= τ+ (2.21) SS(X τ |I) − 2T SS(Xτ |I, X β ) E (t − k)2 where SS(X τ |I) and SS(X τ |I, Xβ , ) are given in Tables 1.2 and 1.1, respectively, T = (T1 , T2 , . . . , Tt ) is the vector of the treatment totals and τ is given by (2.13). Theorem 2.1 For t ≥ 6 and  Z t −k   −1 ρ = t (k − 1) (t − 1)MSE   1 otherwise

if

Z k(t − 1) > MSE t −k

(2.22)

with Z defined by (2.21) and MSE = MS(I |I, X β , Xτ ) being the residual mean square as given in Table 1.1, the combined estimator, c τ , for c τ is uniformly better than the intrablock estimator, c τ , that is, var(c τ ) < var(c τ) For the proof we refer to Shah (1964). Similar estimators for ρ have been developed by Seshadri (1963) and still others by Graybill and Weeks (1959) and Graybill and Deal (1959). We shall comment briefly on the nature of these estimators without going into any detail. Of particular interest is, of course, their relationship to the Yates estimator and the comparison of the performance of all the estimators with regard to estimating differences of treatment effects. We note that the Yates estimator utilizes all b − 1 d.f. between blocks as does the estimator proposed by Graybill and Weeks (1959), which, however, was shown by Seshadri (1963) to be not as good as the Yates estimator. All the other estimators utilize only part of the b − 1 d.f., either t − 1 or b − t. Under certain conditions these estimators have been shown to give rise to estimators for treatment differences uniformly better than the corresponding intrablock estimators. Such a property has not yet been proved for the Yates procedure, but the fact that it uses more information than the other estimators may lead one to suspect that it might actually be better than the estimators mentioned above that enjoy this property. To some extent this aspect was investigated by Shah (1970) along the following lines.

79

SIGNIFICANCE TESTS

Let V be the average variance for treatment comparisons. Then V can be expressed as V =

2σe2 r∗

where r ∗ is defined as the effective number of replications. For example, r ∗ corresponding to the intrablock estimator is given by rI∗ = rE. The largest possible value for r ∗ is that corresponding to the combined estimator when ρ is known. Denote this value by rc∗ . Since usually ρ is not known and hence has to be estimated from the data, some loss of information is incurred; that is, r ∗ ( ρ ) < rc∗ , ∗ where r ( ρ ) is the effective number of replications when an estimator ρ is used instead of ρ in the combined analysis. Shah (1970) shows that for his estimator (given in Theorem 2.1) the difference between rc∗ and r ∗ ( ρ ) is not appreciable for moderate values of ρ. This implies that the loss of information is not great and not much improvement, if any, can be expected from any other estimation procedure, including the Yates procedure. On the other hand we know that for Shah’s estimator r ∗ ( ρ ) > rI∗ when t ≥ 6 (see Theorem 2.1), which is not guaranteed for the Yates procedure. For large values of ρ the difference between rc∗ and rI∗ is fairly small, so that not much is gained by using a combined estimator. The conclusion from this discussion then is that the Shah procedure leads in general to quite satisfactory results and should therefore be used in the combined analysis.

2.6 SIGNIFICANCE TESTS As mentioned in Section 1.13.4 significance tests concerning the treatment effects are performed as approximate F test by substituting ρ for ρ in the coefficient matrix A of (1.60), using any of the previously described methods of estimating ρ. Such tests are mostly conveniently performed by choosing any option in SAS PROC MIXED. An exact test, based, however, on the assumption of normality, for H0 : τ1 = τ2 = · · · = τt against H1 :

not all τi equal

was developed by Cohen and Sackrowitz (1989). We shall give a brief description here, but the reader should refer to the original article for details. The test is based on invariance properties and utilizes information from the intrablock analysis (as described in Sections 2.4.1 and 1.3) and the interblock analysis (as described for the general case in Section 1.7.2) by combining the P values from

80


the two respective independent tests. In order to define the test we establish the following notation: τ = solution to the intrablock normal equations, (2.13) τ ∗ = solution to the interblock normal equations, (1.48), with L = I O = matrix of t − 1 orthonormal contrasts for treatment effects τ U 1 = O U 2 = Oτ ∗ U 1 2 = U 1 U 1 U 2 2 = U 2 U 2 V 1 = U 1/ U 1 V 2 = U 2/ U 2 R = V 1 V 2 S 2 = SS(Error from intrablock analysis) S ∗ = SS(Error from interblock analysis) 2 b t ∗ ∗ Bj − kµ − nij τi = 2

j =1

i=1

λt U 1 2 k r −λ 2 T2 = S ∗ + U 2 2 k a = min(T1 /T2 , 1) T1 = S 2 +

γ = 1/(a + 1) P = P value for testing H0 using intrablock analysis P ∗ = P value for testing H0 using interblock analysis, based on F statistic (b − t)(r − λ) U 2 2 with t − 1 and b − t d.f. 2 k(t − 1) S∗ Z1 = −n P

F∗ =

Z2 = −n P ∗ z = γ Z1 + (1 − γ )Z2

81

SIGNIFICANCE TESTS

Then, for γ = 12 , the P value for the exact test is given by P ∗∗ = γ e−z/γ − (1 − γ )e−z/(1−γ ) /[(2γ − 1)(1 + R)] ! and for γ = 12 by (2z + 1)e−2z /(1 + R). Cohen and Sackrowitz (1989) performed some power simulations and showed that the exact test is more powerful than the usual (approximate) F test. For a more general discussion see also Mathew, Sinha, and Zhou (1993) and Zhou and Mathew (1993). We shall illustrate the test by using an example from Lentner and Bishop (1993). Example 2.2 An experiment is concerned with studying the effect of six diets on weight gain of rabbits using 10 litters of three rabbits each. The data are given in Table 2.1A. The parameters of the BIB design are t = 6, b = 10, k = 3, r = 5, λ = 2. The intrablock analysis using SAS PROC GLM is presented in Table 2.1A yielding F = 3.16 and P = 0.0382. In addition, Table 2.1A gives the SAS solutions to the normal equations, which are then used to compute the set of five orthonormal contrasts based on 

√ −5/ 70

√ −3/ 70

√ −1/ 70

√ 1/ 70

√ 3/ 70

√ 5/ 70

  √ √ √ √ √ √  −1/ 84 −4/ 84 −4/ 84 −1/ 84 5/ 84  5/ 84    √ √ √ √ √ √ −5/ 180 7/ 180 4/ 180 −4/ 180 −7/ 180 5/ 180    √ √ √ √ √ √   1/ 28 −3/ 28 2/ 28 2/ 28 −3/ 28 1/ 28   √ √ √ √ √ √ −1/ 252 5/ 252 −10/ 252 10/ 252 −5/ 252 1/ 252

               

The linear, quadratic, cubic, quartic, and quintic parameter estimates represent the elements of U 1 , yielding U 1 2 = 39.6817. A solution to the interblock normal equations [using L = I in (1.48)] is given in Table 2.1B. The estimates τ ∗ are then used to obtain U 2 and U 2 2 = 537.5284. We also obtain from Table 2.1B F ∗ = 2.23 and P ∗ = 0.2282. Using S 2 = 150.7728 from Table 2.1A and S ∗2 = 577.9133 from Table 2.1B, all other values needed for the test can be computed. With γ = 0.7828 we obtain Z = 2.8768, and finally P ∗∗ = 0.03. This result is in good agreement with the P value of 0.0336 obtained from the combined analysis using SAS PROC MIXED with the REML option (see Table 2.1A).

82


Table 2.1A Data, Intrablock, and Combined Analysis for BIB Design (t = 6, b = 10, r = 5, k = 3, LAMBDA=2) options nodate pageno=1; data rabbit1; input B T Y @@; datalines; 1 6 42.2 1 2 32.6 1 3 35.2 2 3 40.9 2 1 40.1 2 2 38.1 3 3 34.6 3 6 34.3 3 4 37.5 4 1 44.9 4 5 40.8 4 3 43.9 5 5 32.0 5 3 40.9 5 4 37.3 6 2 37.3 6 6 42.8 6 5 40.5 7 4 37.9 7 1 45.2 7 2 40.6 8 1 44.0 8 5 38.5 8 6 51.9 9 4 27.5 9 2 30.6 9 5 20.6 10 6 41.7 10 4 42.3 10 1 37.3 ; run; proc print data=rabbit1; title1 'TABLE 2.1A'; title2 'DATA FOR BIB DESIGN'; title3 '(t=6, b=10, r=5, k=3, LAMBDA=2)'; run; proc glm data=rabbit1; class B T; model Y=B T/solution; estimate 'linear' T -5 -3 -1 1 3 5/divisor=8.3667; estimate 'quad' T 5 -1 -4 -4 -1 5/divisor=9.1652; estimate 'cubic' T -5 7 4 -4 -7 5/divisor=13.4164; estimate 'quartic' T 1 -3 2 2 -3 1/divisor=5.2915; estimate 'quintic' T -1 5 -10 10 -5 1/divisor=15.8745; title2 'INTRA-BLOCK ANALYSIS'; title3 'WITH ORTHONORMAL CONTRASTS'; run; proc mixed data=rabbit1; class B T; model Y=T/solution; random B; lsmeans T; title2 'COMBINED INTRA- AND INTER-BLOCK ANALYSIS'; title3 '(USING METHOD DESCRIBED IN SECTION 1.8)'; run;

83

SIGNIFICANCE TESTS

Table 2.1A (Continued ) Obs

B

T

Y

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

1 1 1 2 2 2 3 3 3 4 4 4 5 5 5 6 6 6 7 7 7 8 8 8 9 9 9 10 10 10

6 2 3 3 1 2 3 6 4 1 5 3 5 3 4 2 6 5 4 1 2 1 5 6 4 2 5 6 4 1

42.2 32.6 35.2 40.9 40.1 38.1 34.6 34.3 37.5 44.9 40.8 43.9 32.0 40.9 37.3 37.3 42.8 40.5 37.9 45.2 40.6 44.0 38.5 51.9 27.5 30.6 20.6 41.7 42.3 37.3

The GLM Procedure Class Level Information Class B T

Levels 10 6

Values 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6

84


Table 2.1A (Continued ) Number of observations 30 Dependent Variable: Y

Source

DF

Sum of Squares

Mean Square

Model Error Corrected Total

14 15

889.113889 150.772778

63.508135 10.051519

29

1039.886667

F Value

Pr > F

6.32

0.0005

R-Square

Coeff Var

Root MSE

Y Mean

0.855010

8.241975

3.170413

38.46667

Source B T Source B T

DF

Type I SS

Mean Square

F Value

Pr > F

9 5

730.3866667 158.7272222

81.1540741 31.7454444

8.07 3.16

0.0002 0.0382

DF

Type III SS

Mean Square

F Value

Pr > F

9 5

595.7352222 158.7272222

66.1928025 31.7454444

6.59 3.16

0.0008 0.0382

Parameter linear quad cubic quartic quintic

0.68326421 2.35674071 3.14664639 4.74975590 1.09504761

Parameter Intercept B B

Standard Error

t Value

Pr > |t|

1.58518760 1.58519809 1.58520742 1.58520728 1.58520728

0.43 1.49 1.99 3.00 0.69

0.6726 0.1578 0.0657 0.0090 0.5002

Estimate

Estimate

1 2

42.61111111 B -3.29722222 B 0.83611111 B

Standard Error

t Value

Pr > |t|

2.24182052 2.79604144 2.79604144

19.01 -1.18 0.30

|t|

2.69433295 2.79604144 2.79604144

-1.89 1.97 -0.35

0.0778 0.0681 0.7278

2.79604144 2.69433295 2.69433295 2.79604144 . 2.24182052 2.24182052 2.24182052 2.24182052 2.24182052 .

0.76 0.92 2.28 -3.85 . -1.47 -2.25 -1.29 -1.44 -3.80 .

0.4619 0.3718 0.0380 0.0016 . 0.1617 0.0400 0.2154 0.1698 0.0017 .

NOTE: The X'X matrix has been found to be singular, and a generalized inverse was used to solve the normal equations. Terms whose estimates are followed by the letter 'B' are not uniquely estimable.

COMBINED INTRA- AND INTERBLOCK ANALYSIS (USING METHOD DESCRIBED IN SECTION 1.8) The Mixed Procedure Model Information Data Set Dependent Variable Covariance Structure Estimation Method Residual Variance Method Fixed Effects SE Method Degrees of Freedom Method

WORK.RABBIT1 Y Variance Components REML Profile Model-Based Containment

Class Level Information Class B T

Levels 10 6

Values 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6

86


Table 2.1A (Continued ) Dimensions Covariance Parameters Columns in X Columns in Z Subjects Max Obs Per Subject Observations Used Observations Not Used Total Observations

2 7 10 1 30 30 0 30

Iteration History Iteration

Evaluations

-2 Res Log Like

Criterion

0 1 2 3

1 2 1 1

160.26213715 150.36288495 150.35691057 150.35688183

0.00010765 0.00000054 0.00000000

Convergence criteria met. Covariance Parameter Estimates Cov Parm

Estimate

B Residual

21.6953 10.0840 Fit Statistics

-2 Res Log Likelihood AIC (smaller is better) AICC (smaller is better) BIC (smaller is better)

150.4 154.4 154.9 155.0

Solution for Fixed Effects

Effect

T

Estimate

Standard Error

DF

t Value

Pr > |t|

Intercept T

1

42.3454 -2.8100

2.1303 2.2087

9 15

19.88 -1.27

|t|

T T T T T

2 3 4 5 6

-5.3172 -2.9941 -3.6952 -8.4560 0

2.2087 2.2087 2.2087 2.2087 .

15 15 15 15 .

-2.41 -1.36 -1.67 -3.83 .

0.0294 0.1953 0.1150 0.0016 .

Type 3 Tests of Fixed Effects Effect

Num DF

Den DF

F Value

Pr > F

5

15

3.28

0.0336

T

Least Squares Means

Effect

T

Estimate

Standard Error

DF

t Value

Pr > |t|

T T T T T T

1 2 3 4 5 6

39.5354 37.0282 39.3513 38.6502 33.8894 42.3454

2.1303 2.1303 2.1303 2.1303 2.1303 2.1303

15 15 15 15 15 15

18.56 17.38 18.47 18.14 15.91 19.88

2 associate classes. The eigenvalues given in (4.33) together with their multiplicities for u = 1, 2 can be used to obtain the efficiency factor of a given design (see Section 1.11). We shall also make use of them in Section 4.7 in connection with the combined analysis of PBIB designs.

4.5 ANALYSIS OF PBIB DESIGNS 4.5.1 Intrablock Analysis The analysis of PBIB(m) designs can be derived easily from the general analysis of incomplete block designs as presented in Section 1.3. In order to solve the RNE [see (1.8)] C τ =Q where C is given in (4.3) and Q is the vector of adjusted treatment totals, we shall find a generalized inverse C − for C satisfying (1.17); that is, 1 CC − = I − II t by utilizing (4.11) and (4.12) as outlined previously. This leads to the following system of m + 1 equations in m + 1 unknowns g0 , g1 , . . . , gm : m m

k cu puv gv = 1 −

u=0 v=0

=−

1 t

1 t

for

k=0

for k = 1, 2, . . . , m

(4.34)

132

PARTIALLY BALANCED INCOMPLETE BLOCK DESIGNS

Since m m

k cu puv =

u=0 v=0

m

c u nu = 0

u=0

it follows that this system of equations is not of full rank. Therefore any m of the equations in (4.34) can be

taken and solved with an additional convenient restriction like gu = 0 or nu gu = 0, or for some v, gv = 0. We shall illustrate this procedure for a PBIB(2) design. The system (4.34) is then of the form c 0 g 0 + c 1 n1 g 1 + c 2 n2 g 2 = 1 − 1 t 1 2 2 2 2 + c2 p12 )g1 + (c0 + c1 p12 + c2 p22 )g2 = − c2 g0 + (c1 p11 t

1 1 1 1 + c2 p12 )g1 + (c1 p12 + c2 p22 )g2 = − c1 g0 + (c0 + c1 p11

1 t (4.35)

with c0 =

r(k − 1) k

c1 = −

λ1 k

c2 = −

λ2 k

Letting g0 = 0 and omitting the first equation then yields 1 t 1 2 2 2 2 + c2 p12 )g1 + (c0 + c1 p12 + c2 p22 )g2 = − (c1 p11 t

1 1 1 1 + c2 p12 )g1 + (c1 p12 + c2 p22 )g2 = − (c0 + c1 p11

(4.36)

Example 4.2 Consider the following PBIB(2) design (SR 36 in Clatworthy, 1973) with parameters t = 8, r = 4, k = 4, b = 8, λ1 = 0, λ2 = 2: Block 1 2 3 4 5 6 7 8

Treatments 1, 5, 2, 6, 3, 7, 4, 8,

2, 6, 7, 3, 8, 4, 1, 5,

3, 7, 8, 4, 1, 5, 6, 2,

4 8 1 5 6 2 7 3

133

ANALYSIS OF PBIB DESIGNS

which has the following association scheme: 0th Associate

1st Associates

1 2 3 4 5 6 7 8

5 6 7 8 1 2 3 4

2nd Associates 2, 1, 1, 1, 2, 1, 1, 1,

3, 3, 2, 2, 3, 3, 2, 2,

4, 4, 4, 3, 4, 4, 4, 3,

6, 5, 5, 5, 6, 5, 5, 5,

7, 7, 6, 6, 7, 7, 6, 6,

8 8 8 7 8 8 8 7

giving n1 = 1, n2 = 6, and  1 P0 =  1



  0 1 0 P 1 = 1 0 0 0 0 6

 6

  0 0 1 P 2 = 0 0 1 1 1 4

Hence c0 = 3, c1 = 0, c2 = − 12 , and (4.36) becomes 3g1 − 3g2 = − 18 − 12 g1 + g2 = − 18 yielding g1 = −0.3333 and g2 = −0.2917. Having solved (4.34) or its reduced form for g0 , g1 , . . . , gm , we obtain the intrablock estimates of treatment effects as τ=

m

(4.37)

gv B v Q

v=0

or, equivalently, τi = g0 Qi + g1

j S(i,1)

Qj + g2

Qk + · · · + gm

k S(i,2)

Q

(4.38)

S(i,m)

(i = 1, 2, . . . , t) where S(i, u) is, as defined earlier, the set of uth associates of treatment i. It follows further from (4.37) and the definition of the association matrices that var( τi − τi ) = 2(g0 − gu )σe2

(4.39)

134


if treatments i and i are uth associates. There are, therefore, m types of comparisons which, however, in certain cases may not all be distinct. Correspondingly, there are also at most m types of efficiency factors, Eu . Since var( τi − τi ) =

2σe2 r

for a CRD or RCBD, we have Eu =

1 r(g0 − gu )

(4.40)

(u = 1, 2, . . . , m) (see Section 1.11). The overall efficiency factor of a PBIB(m) design relative to the CRD or the RCBD is then given by E=

m 1 nu E u t −1

(4.41)

u=1

For the PBIB(2) design discussed in Example 4.2, we find (since g0 = 0) var( τi − τi ) = .6666σe2 for 1st associates and τi ) = .5834σe2 var( τi − for 2nd associates. Also, from (4.40) and (4.41) E1 = −

1 = .750 rg1

E2 = −

1 = .875 rg2

E = 17 (E1 + 6E2 ) = .842

4.5.2 Combined Analysis As developed in Section 1.8.3, the set of normal equations for the combined analysis is of the form A τ =P

(4.42)

with A = rI −

1 − ρ −1 NN k

(4.43)

135

ANALYSIS OF PBIB DESIGNS

and P =T −

1 − ρ −1 NB k

(4.44)

In order to obtain τ we use the same type of argument as in Section 4.2.1, that is, make use of the special form of A as expressed in terms of association matrices; that is, A=

m

(4.45)

au B u

u=0

with

1 − ρ −1 a0 = r 1 − k au = −

1 − ρ −1 λu k

(4.46)

(u = 1, 2, . . . , m)

(4.47)

Writing A−1 as A−1 =

m

auB u

u=0

we obtain, equivalent to the set of equations (4.34) for the intrablock analysis, the following system of equations: m m m −1 v k AA = au B u a Bv = au a v puv Bk = I (4.48) u=0

uv k=0

v=0

or equivalently, m m

k v au puv a = δ0k

(k = 0, 1, . . . , m)

(4.49)

u=0 v=0

where δ0k is the Kronecker symbol. This set of equations is of full rank and hence we have (m + 1) equations in (m + 1) unknowns. It follows then that τ i = a 0 Pi + a 1 Pj + a 2 Pk + · · · + a m P (4.50) j S(i,1)

k S(i,2)

S(i,m)

where the Pu (u = 1, 2, . . . , t) are the elements of P in (4.44).

136


For any two treatments that are uth associates we have consequently var τi − τ i = 2(a 0 − a u )σe2

(4.51)

We shall illustrate the procedure by using Example 4.2 given in Section 4.4.1. Example 4.2 (Continued) 0, λ2 = 2.

We have t = 8, r = λ0 = 4, k = 4, b = 8, λ1 =

 1  P0 =  1

   6

  0 1 0   P 1 = 1 0 0 0 0 6

  0 0 1   P 2 = 0 0 1 1 1 4 Then, from (4.46) and (4.47), 1 − ρ −1 a0 = 4 1 − = 3 + ρ −1 4 a1 = 0 a2 = −(1 − ρ −1 ) 12 and hence Eqs. (4.49) are given as (3 + ρ −1 ) a 0

− 3(1 − ρ −1 ) a 2 = 1 (3 + ρ −1 ) a 1

− 12 (1 − ρ −1 ) a 0 −

1 2

− 3(1 − ρ −1 ) a 2 = 0

(1 − ρ −1 ) a 1 + [3 + ρ −1 − 2 (1 − ρ −1 )] a 2 = 0

There remains, of course, the problem of estimating ρ or ρ −1 . Substituting such an estimator in the equations above, or more generally equations (4.49), will then yield solutions for a 0 , a 1 , a 2 which we denote by a 0 , a 1 , a2 and which, in turn, are used in Eqs. (4.50) and (4.51). One such estimator is the Yates estimator as described in Section 1.10.1. For certain types of PBIB(2) designs we shall present in Section 4.7 other estimators for ρ similar to the one given for BIB designs (see Section 2.5), which will lead to uniformly better estimators (than the intrablock estimators) for treatment contrasts under certain conditions.

CLASSIFICATION OF PBIB DESIGNS

137

4.6 CLASSIFICATION OF PBIB DESIGNS Since their introduction by Bose and Nair (1939) and Bose and Shimamoto (1952) different types of PBIB designs have evolved. These can be classified according to the number of associate classes and type of association scheme. (We mention here parenthetically that the association scheme and the parameters of a PBIB design do not lead immediately to the actual experimental layout or field plan; see also Chapter 5). In this section we shall give a brief survey (without being complete) of types of PBIB designs. The designs perhaps most often used are PBIB(2) designs. The reason, of course, is that they constitute the simplest extension of the BIB design. They have been tabulated (for certain ranges of parameters) extensively by Bose, Clatworthy, and Shrikhande (1954) and Clatworthy (1956, 1973) and are classified according to which of the following association schemes (to be explained in the following sections) is being used: 1. 2. 3. 4.

Group-divisible PBIB(2) designs Triangular PBIB(2) designs Latin square type PBIB(2) designs Cyclic PBIB(2) designs

We then mention the association schemes and parameters for some PBIB(3) designs: 1. 2. 3. 4.

Rectangular PBIB(3) designs Generalized group-divisible PBIB(3) designs Generalized triangular PBIB(3) designs Cubic PBIB(3) designs

Finally, we discuss some rather general classes of PBIB designs with m > 3 associate classes: 1. 2. 3. 4.

Extended group-divisible PBIB designs Hypercubic PBIB designs Right-angular PBIB(4) designs Cyclic PBIB designs

4.6.1 Group-Divisible (GD) PBIB(2) Designs Suppose that t can be written as t = t1 t2 and that the treatments are divided into t1 groups of t2 treatments each by arranging them in a rectangular array of t1 rows and t2 columns (t1 , t2 > 1), the rows constituting the t1 groups. Then two treatments are called first associates if they are in the same group and they are called second associates otherwise.

138


This association scheme implies n1 = t2 − 1 and n2 = (t1 − 1)t2 . Further 

 0 1  P 1 = 1 t2 − 2

0

0

(t1 − 1)t2

0

 0  P 2 = 0

 

0

0

1



 t2 − 1  1 t2 − 1 (t1 − 2)t2 0

Example 4.3 Let t = 8, t1 = 2, t2 = 4. Association scheme: 1 3 5 7 2 4 6 8 0th Associate

1st Associates

2nd Associates

1 3, 5, 7 2, 4, 6, 8 2 4, 6, 8 1, 3, 5, 7 3 1, 5, 7 2, 4, 6, 8 4 2, 6, 8 1, 3, 5, 7 5 1, 3, 7 2, 4, 6, 8 6 2, 4, 8 1, 3, 5, 7 7 1, 3, 5 2, 4, 6, 8 8 2, 4, 6 1, 3, 5, 7 The factorization of t is, of course, not always unique. For Example 4.3 we could just as well have t1 = 4, t2 = 2, in which case the rectangular array is 1 2 3 4

5 6 7 8

and hence n1 = 1, n2 = 6. Bose and Connor (1952) have shown that for GD-PBIB(2) designs the following inequalities hold: r ≥ λ1

rk − λ2 t ≥ 0

Accordingly, they have divided the GD designs further into three subclasses: 1. Singular if r = λ1 2. Semiregular if r > λ1 , rk − λ2 t = 0 3. Regular if r > λ1 , rk − λ2 t > 0

139


4.6.2 Triangular PBIB(2) Designs Suppose that t can be written as t = q(q − 1)/2 and that the treatments are arranged in triangular form above the main diagonal of a q × q array and repeated symmetrically below the main diagonal leaving the main diagonal blank. Then two treatments are called first associates if they lie in the same row (or column) of the q × q array and are second associates otherwise. This association scheme implies n1 = 2q − 4

n2 = (q − 2)(q − 3)/2

 0 1  P 1 = 1 q − 2



0

 

q −3

0 q − 3 (q − 3)(q − 4)/2  0  P 2 = 1

0

1

4

2q − 8

  

1 2q − 8 (q − 4)(q − 5)/2 Example 4.4 Let t = 10, q = 5. Association scheme: ∗ 1 2 3 4 0th Associate 1 2 3 4 5 6 7 8 9 10

1 ∗ 5 6 7

2 3 4 5 6 7 ∗ 8 9 8 ∗ 10 9 10 ∗

1st Associates 2, 1, 1, 1, 1, 1, 1, 2, 2, 3,

3, 3, 2, 2, 2, 3, 4, 3, 4, 4,

4, 4, 4, 3, 6, 5, 5, 5, 5, 6,

5, 5, 6, 7, 7, 7, 6, 6, 7, 7,

6, 8, 8, 9, 8, 8, 9, 9, 8, 8,

2nd Associates 7 9 10 10 9 10 10 10 10 9

8, 6, 5, 5, 3, 2, 2, 1, 1, 1,

9, 7, 7, 6, 4, 4, 3, 4, 3, 2,

10 10 9 8 10 9 8 7 6 5

140


4.6.3 Latin Square Type PBIB(2) Designs Suppose that t can be written as t = q 2 and that the treatments are arranged in a square array of size q. For a Latin square type PBIB(2) design with two constraints [L2 -PBIB(2) design], two treatments are called first associates if they occur together in the same row or the same column of the square array and are second associates otherwise. For the Latin square type design with i constraints (i > 2) [Li -PBIB(2) design] the association scheme is used in conjunction with i − 2 mutually orthogonal Latin squares (if they exist) (see I.10.6.2). Then two treatments are called first associates if they both occur in the same row or column or correspond to the same letter of the superimposed Latin squares. Hence we obtain for the Li -PBIB(2) design n1 = i(q − 1) n2 = (q − 1)(q − i + 1)   0 1 0 P 1 = 1 i 2 − 3i + q (i − 1)(q − i + 1)  0 (i − 1)(q − i + 1) (q − i)(q − i + 1)   0 0 1  i(q − 1) P 2 = 0 i(i − 1) 1 i(q − i) (q − i)2 + i − 2 Example 4.5 Let t = 9, q = 3. L2 -association scheme: 1 2 3 4 5 6 7 8 9 0th Associate 1 2 3 4 5 6 7 8 9

1st Associates 2, 1, 1, 1, 2, 3, 1, 2, 3,

3, 3, 2, 5, 4, 4, 4, 5, 6,

4, 5, 6, 6, 6, 5, 8, 7, 7,

7 8 9 7 8 9 9 9 8

2nd Associates 5, 4, 4, 2, 1, 1, 2, 1, 1,

6, 6, 5, 3, 3, 2, 3, 3, 2,

8, 7, 7, 8, 7, 7, 5, 4, 4,

9 9 8 9 9 8 6 6 5

141


Example 4.6 Let t = 9, q = 3. L3 -association scheme: 1A 2B 3C 4C 5A 6B 7B 8C 9A 0th Associate 1 2 3 4 5 6 7 8 9

1st Associates 2, 1, 1, 1, 1, 2, 1, 2, 1,

3, 3, 2, 3, 2, 3, 2, 3, 3,

4, 5, 4, 5, 4, 4, 4, 4, 5,

5, 6, 6, 6, 6, 5, 6, 5, 6,

7, 7, 8, 7, 8, 7, 8, 7, 7,

2nd Associates 9 8 9 8 9 9 9 9 8

6, 4, 5, 2, 3, 1, 3, 1, 2,

8 9 7 9 7 8 5 6 4

Other examples are mentioned in Section 18.7. 4.6.4 Cyclic PBIB(2) Designs Suppose the treatments are denoted by 0, 1, 2, . . . , t − 1. Then the first associates of treatment i are i + d1 , i + d2 , . . . , i + dn1 (mod t) where the dj are integers satisfying the following conditions: 1. The dj are all different with 0 < dj < t for each j . 2. Among the n1 (n1 − 1) differences dj − dj (mod t) each of the integers d1 , d2 , . . . , dn1 occurs α times, and each of the integers e1 , e2 , . . . , en2 occurs β times, where d1 , d2 , . . . , dn1 , e1 , e2 , . . . , en2 are all the different integers 1, 2, . . . , t − 1. The second associates of treatment i obviously are i + e1 , i + e2 , . . . , i + en2 (mod t). We then have, as a consequence of condition 2, n1 α + n2 β = n1 (n1 − 1) and

  0 1 0 α n1 − α − 1  P 1 = 1 0 n1 − α − 1 n2 − n1 + α + 1

142


 1  n1 − β n2 − n1 + β − 1

 0 0 P 2 = 0 β 0 n1 − β

Example 4.7 Let t = 5, d1 = 2, d2 = 3, α = 0, β = 1. With n1 = 2, the possible differences dj − dj are d1 − d2 = −1 = 4 and d2 − d1 = 1; that is, d1 and d2 occur α = 0 times among these differences and e1 = 1, e2 = 4 occur β = 1 time. Hence condition 2 is satisfied. We then have 0th Associate 0 1 2 3 4

1st Associates 2, 3, 4, 0, 1,

3 4 0 1 2

2nd Associates 1, 0, 1, 2, 0,

4 2 3 4 3

4.6.5 Rectangular PBIB(3) Designs This association scheme was first proposed by Vartak (1959). Suppose that t = t1 t2 and that the treatments are arranged in a rectangular array with t1 rows and t2 columns. Two treatments are said to be first associates if they occur together in the same row of the array; they are said to be second associates if they occur together in the same column; they are third associates otherwise. Hence n1 = t2 − 1,

n2 = t1 − 1

n3 = (t1 − 1)(t2 − 1)

and   0 1 0 0  1 t2 − 2 0 0  P1 =   0 0 0 t1 − 1 0 0 t1 − 1 (t1 − 1)(t2 − 2)   0 0 1 0  0 0 0 t2 − 1  P2 =   1 0 t1 − 2 0 0 t2 − 1 0 (t1 − 2)(t2 − 1)  0 0 P3 =  0

 0 0 1  0 1 t2 − 2   1 0 t1 − 2 t2 − 2 t1 − 2 (t1 − 2)(t2 − 2)

143


Example 4.8 Let t = 6, t1 = 2, t2 = 3. Association scheme:

0th Associate 1 2 3 4 5 6

1 3 5 2 4 6 1st Associates 2nd Associates 3, 4, 1, 2, 1, 2,

5 6 5 6 3 4

3rd Associates

2 1 4 3 6 5

4, 3, 2, 1, 2, 1,

6 5 6 5 4 3

4.6.6 Generalized Group-Divisible (GGD) PBIB(3) Designs In a more general way this association scheme was proposed by Roy (1953–1954) and described in more detail by Raghavarao (1960). Suppose that t = t1 t2 t3 and that the treatments are arranged in a threedimensional array with t1 rows, t2 columns, and t3 layers. Two treatments are said to be first associates if they occur together in the same row and column, but in different layers; they are said to be second associates if they occur in the same row but in different columns (and the same or different layers); they are third associates otherwise. Hence n1 = t3 − 1

n2 = (t2 − 1)t3

n3 = (t1 − 1)t2 t3

and  0 1 0 1 t − 2 0 3  P1 =  0 0 (t2 − 1)t3 0

0

0

 0 0 1 0 0 t3 − 1  P2 =  1 t3 − 1 (t2 − 2)t3 0  0 0  P3 =  0

0 0 0

   

(t1 − 1)t2 t3 0 0 0

0

0

(t1 − 1)t2 t3

0

0

1

0

0

0

0

1 t3 − 1 (t2 − 1)t3



     

   (t2 − 1)t3  t3 − 1

(t1 − 2)t2 t3

144


Example 4.9 Let t = 12, t1 = 2, t2 = 3, t3 = 2. Association scheme: Layer 1 1 2

3 4

0th Associate

1st Associates

1 2 3 4 5 6 7 8 9 10 11 12

7 8 9 10 11 12 1 2 3 4 5 6

Layer 2

5 6

7 8

9 10

2nd Associates 3, 4, 1, 2, 1, 2, 3, 4, 1, 2, 1, 2,

5, 6, 5, 6, 3, 4, 5, 6, 5, 6, 3, 4,

9, 10, 7, 8, 7, 8, 9, 10, 7, 8, 7, 8,

11 12 11 12 9 10 11 12 11 12 11 10

11 12 3rd Associates 2, 1, 2, 1, 2, 1, 2, 1, 2, 1, 2, 1,

4, 3, 4, 3, 4, 3, 4, 3, 4, 3, 4, 3,

6, 5, 6, 5, 6, 5, 6, 5, 6, 5, 6, 5,

8, 7, 8, 7, 8, 7, 8, 7, 8, 7, 8, 7,

10, 9, 10, 9, 10, 9, 10, 9, 10, 9, 10, 9,

12 11 12 11 12 11 12 11 12 11 12 11

4.6.7 Generalized Triangular PBIB(3) Designs John (1966) has shown that the triangular association scheme discussed in Section 4.5.2 can be described equivalently by representing the treatments by ordered pairs (x, y) with 1 ≤ x < y ≤ q and calling two treatments first associates if they have one integer in common, second associates otherwise. This can then be generalized (John, 1966) for the case that t = q(q − 1)(q − 2)/6 (q > 3) and that the treatments are represented by ordered triplets (x, y, z) with 1 ≤ x < y < z ≤ q. Two treatments are said to be first associates if they have two integers in common, second associates if they have one integer in common but not two, and third associates otherwise. Hence n1 = 3(q − 3)

n2 = 3(q − 3)(q − 4)/2

n3 = (q − 3)(q − 4)(q − 5)/6

and  0 1  1 q−2 P1 =  0 2(q − 4)  0

0

0

0

2(q − 4)

0

(q − 4)2

(q − 4)(q − 5)/2

(q − 4)(q − 5)/2 (q − 4)(q − 5)(q − 6)/6

     

145


 0 0 1 1 4 2(q − 4)  P2 =  1 2(q − 4) (q + 2)(q − 5)/2 0



0

   

q−5 (q − 5)(q − 6)

q −5

(q − 5)(q − 6)

0

0

1

0

9

3(q − 6)

9

9(q − 6)

3(q − 6)(q − 7)/2

 0 0  P3 =  0

(q − 5)(q − 6)(q − 7)/6     

1 3(q − 6) 3(q − 6)(q − 7)/2 (q − 6)(q − 7)(q − 8)/6

Example 4.10 Let t = 20, q = 6. The treatments are represented as 1 ≡ (1, 2, 3)

2 ≡ (1, 2, 4) 5 ≡ (1, 3, 4)

3 ≡ (1, 2, 5) 6 ≡ (1, 3, 5) 8 ≡ (1, 4, 5)

11 ≡ (2, 3, 4)

12 ≡ (2, 3, 5) 14 ≡ (2, 4, 5)

4 ≡ (1, 2, 6) 7 ≡ (1, 3, 6) 9 ≡ (1, 4, 6) 10 ≡ (1, 5, 6) 13 ≡ (2, 3, 6) 15 ≡ (2, 4, 6) 16 ≡ (2, 5, 6) 18 ≡ (3, 4, 6) 19 ≡ (3, 5, 6) 20 ≡ (4, 5, 6)

17 ≡ (3, 4, 5)

and the association scheme follows: 0th Associate 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

1st Associates 2, 1, 1, 1, 1, 1, 1, 2, 2, 3, 1, 1, 1, 2, 2, 3, 5, 5, 6, 8,

3, 3, 2, 2, 2, 3, 4, 3, 4, 4, 2, 3, 4, 3, 4, 4, 6, 7, 7, 9,

4, 4, 4, 3, 6, 5, 5, 5, 5, 6, 5, 6, 7, 8, 9, 10, 8, 9, 10, 10,

5, 5, 6, 7, 7, 7, 6, 6, 7, 7, 12, 11, 11, 11, 11, 12, 11, 11, 12, 14,

6, 8, 8, 9, 8, 8, 9, 9, 8, 8, 13, 13, 12, 12, 13, 13, 12, 13, 13, 15,

7, 9, 10, 10, 9, 10, 10, 10, 10, 9, 14, 14, 15, 15, 14, 14, 14, 15, 16, 16,

11, 11, 12, 13, 11, 12, 13, 14, 15, 16, 15, 16, 16, 16, 16, 15, 18, 17, 17, 17,

2nd Associates 12, 14, 14, 15, 17, 17, 18, 17, 18, 19, 17, 17, 18, 17, 18, 19, 19, 19, 18, 18,

13 15 16 16 18 19 19 20 20 20 18 19 19 20 20 20 20 20 20 19

8, 6, 5, 5, 3, 2, 2, 1, 1, 1, 3, 2, 2, 1, 1, 1, 1, 1, 1, 2,

9, 7, 7, 6, 4, 4, 3, 4, 3, 2, 4, 4, 3, 4, 3, 2, 2, 2, 3, 3,

10, 10, 9, 8, 10, 9, 8, 7, 6, 5, 6, 5, 5, 5, 5, 6, 3, 4, 4, 4,

14, 12, 11, 11, 12, 11, 11, 11, 11, 12, 7, 7, 6, 6, 7, 7, 7, 6, 5, 5,

15, 13, 13, 12, 13, 13, 12, 12, 13, 13, 8, 8, 9, 9, 8, 8, 9, 8, 8, 6,

16, 16, 15, 14, 14, 14, 15, 15, 14, 14, 9, 10, 10, 10, 10, 9, 10, 10, 9, 7,

17, 17, 17, 18, 15, 16, 16, 16, 16, 15, 16, 15, 14, 13, 12, 11, 13, 12, 11, 11,

3rd Associates 18, 18, 19, 19, 19, 18, 17, 18, 17, 17, 19, 18, 17, 18, 17, 17, 15, 14, 13, 12,

19 20 20 20 20 20 20 19 19 18 20 20 20 19 19 18 16 16 15 13

20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1

146


4.6.8 Cubic PBIB(3) Designs This association scheme, described by Raghavarao and Chandrasekhararao (1964), is an extension of the L2 -association scheme discussed in Section 4.5.3, which can be described equivalently as follows: Denote the t = q 2 treatments by pairs (x, y) with 1 ≤ x, y ≤ q. Define the distance δ between two treatments (x, y) and (x , y ) as the number of nonzero elements in (x − x , y − y ). For the L2 -association scheme then two treatments are said to be first associates if δ = 1, and second associates if δ = 2. For the cubic association scheme we have t = q 3 , each treatment being represented by a triplet (x, y, z) with 1 ≤ x, y, z ≤ q. If the distance δ between any two treatments (x, y, z) and (x , y , z ) is the number of nonzero elements in (x − x , y − y , z − z ), then two treatments are said to be first, second, or third associates if δ = 1, 2, or 3, respectively. This can be interpreted geometrically by realizing that the above representation of the treatments corresponds to an arrangement in a cube of side q. Two treatments are then first associates if they are lying on the same axis, second associates if they are lying on the same plane (but not on the same axis), and third associates otherwise. The design given in Example 4.1 has this association scheme with q = 2. The cubic association scheme implies immediately n1 = 3(q − 1)

n2 = 3(q − 1)2

n3 = (q − 1)3

and  0 1 0  q −2 2(q − 1) 1 P1 =  0 2(q − 1) 2(q − 1)(q − 2)  0

0

    2 (q − 1)  2 (q − 1) (q − 2) 0

(q − 1)2

 0 0 1  2 2(q − 2) 0 P2 =  1 2(q − 2) 2(q − 1) + (q − 2)2  0  0  0 P3 =  0 

q −1

2(q − 1)(q − 2)

0

0

0

3

3

6(q − 2)

1 3(q − 2) 3(q − 2)2



0

1

0



   2(q − 1)(q − 2)  2 (q − 1)(q − 2) q−1



 3(q − 2)   3(q − 2)2   3 (q − 2)

147


Example 4.11 Let t = 8, q = 2. The treatments are represented as 1 ≡ (1, 1, 1) 3 ≡ (1, 2, 1) 5 ≡ (2, 1, 1) 7 ≡ (2, 2, 1)

2 ≡ (1, 1, 2) 4 ≡ (1, 2, 2) 6 ≡ (2, 1, 2) 8 ≡ (2, 2, 2)

Association scheme: 0th Associate 1 2 3 4 5 6 7 8

1st Associates 2, 1, 1, 2, 1, 2, 3, 4,

3, 4, 4, 3, 6, 5, 5, 6,

5 6 7 8 7 8 8 7

2nd Associates 4, 3, 2, 1, 2, 1, 1, 2,

6, 5, 5, 6, 3, 4, 4, 3,

3rd Associates

7 8 8 7 8 7 6 5

8 7 6 5 4 3 2 1

Other examples are mentioned in Section 18.7. 4.6.9 Extended Group-Divisible (EGD) PBIB Designs These designs, which we shall abbreviate as EGD/(2ν − 1)-PBIB, were first introduced by Hinkelmann and Kempthorne (1963) as an extension of the rectangular PBIB(3) design of Vartak (1959), and described in more detail by Hinkelmann (1964). The association scheme for these designs is as follows: Let there be t = t1 t2 , . . . , tν treatments denoted by (i1 , i2 , . . . , iν ), where i = 1, 2, . . . , t and = 1, 2, . . . , ν, and i is called the th component of the treatment. Two treatments are said to be γ th associates, γ = (γ1 , γ2 , . . . , γν ) and γ = 0 or 1( = 1, 2, . . . , ν), if the treatments differ only in the components that correspond to the unity components of γ . We then have m = 2ν − 1 associate classes as each component in γ takes on the possible values 0, 1, and γ = (0, 0, . . . , 0) represents the trivial association corresponding to the 0th associate class of other PBIB designs. The number of associates in the γ th associate class is denoted by n(γ ) and n(γ ) = n(γ1 , γ2 , . . . , γν ) =

ν

(t−1 )γ

(4.52)

=1

If we write the associate classes in lexicographic order, for example, for ν = 3: 000, 001, 010, 011, 100, 101, 110, 111, and if we denote by

148


P 0(t )

P 1(t )

1 0 = 0 t − 1 0 1 = 1 t − 2

the P matrices for a BIB design with t treatments, then we can write the P matrices for the EGD/(2ν − 1)-PBIB design recursively as the following Kronecker product: P γ1 γ2 ...γν−1 0 = P γ1 γ2 ...γν−1 × P 0(tν )

(4.53)

P γ1 γ2 ...γν−1 1 = P γ1 γ2 ...γν−1 × P 1(tν )

(4.54)

where P γ1 γ2 ...γν−1 are the P matrices for an EGD/(2ν−1 − 1) design with t = t1 t2 , . . . , tν−1 treatments. In particular then for ν = 2:   P 00 =  

t2 − 1 ∅

 P 01

0

P 11

∅ 0 t1 − 1 1

 ∅  = 0   1 0 t1 − 2 0 t2 − 1 0 

(t1 − 1)(t2 − 1) 

t2 − 2 ∅

P 10

t1 − 1

  

1

  = 1  



∅

1

0

 ∅  = 1   0 1 0 1 t2 − 2 t1 − 2

t1 − 1 (t1 − 1)(t2 − 2) 0 t2 − 1 0 (t1 − 2)(t2 − 1) 1 t2 − 2 t1 − 2 (t1 − 2)(t2 − 2)

               

which are, apart from the order of rows and columns, the same as those given for the rectangular PBIB(3) design (see Section 4.5.5).

149


Example 4.12 Let t = 36, t1 = t2 = 3, t3 = 4. The treatments are 1 ≡ (1, 1, 1) 5 ≡ (1, 2, 1) 9 ≡ (1, 3, 1) 13 ≡ (2, 1, 1) 17 ≡ (2, 2, 1) 21 ≡ (2, 3, 1) 25 ≡ (3, 1, 1) 29 ≡ (3, 2, 1) 33 ≡ (3, 3, 1)

2 ≡ (1, 1, 2) 6 ≡ (1, 2, 2) 10 ≡ (1, 3, 2) 14 ≡ (2, 1, 2) 18 ≡ (2, 2, 2) 22 ≡ (2, 3, 2) 26 ≡ (3, 1, 2) 30 ≡ (3, 2, 2) 34 ≡ (3, 3, 2)

3 ≡ (1, 1, 3) 7 ≡ (1, 2, 3) 11 ≡ (1, 3, 3) 15 ≡ (2, 1, 3) 19 ≡ (2, 2, 3) 23 ≡ (2, 3, 3) 27 ≡ (3, 1, 3) 31 ≡ (3, 2, 3) 35 ≡ (3, 3, 3)

4 ≡ (1, 1, 4) 8 ≡ (1, 2, 4) 12 ≡ (1, 3, 4) 16 ≡ (2, 1, 4) 20 ≡ (2, 2, 4) 24 ≡ (2, 3, 4) 28 ≡ (3, 1, 4) 32 ≡ (3, 2, 4) 36 ≡ (3, 3, 4)

The associate classes and the numbers of associates are 0 ≡ (0, 0, 0) 1 ≡ (0, 0, 1) 2 ≡ (0, 1, 0) 3 ≡ (0, 1, 1) 4 ≡ (1, 0, 0) 5 ≡ (1, 0, 1) 6 ≡ (1, 1, 0) 7 ≡ (1, 1, 1)

n(0, 0, 0) ≡ 1 n(0, 0, 1) ≡ 3 n(0, 1, 0) ≡ 2 n(0, 1, 1) ≡ 6 n(1, 0, 0) ≡ 2 n(1, 0, 1) ≡ 6 n(1, 1, 0) ≡ 4 n(1, 1, 1) ≡ 12

and the (partial) association scheme is given in Table 4.1

4.6.10 Hypercubic PBIB Designs These designs, first indicated by Shah (1958) and more formally defined by Kusumoto (1965), are extensions of cubic PBIB designs (see Section 4.5.8). Suppose we have t = q ν treatments, which are again denoted by (i1 , i2 , . . . , iν ) where i = 1, 2, . . . , q and = 1, 2, . . . , ν. The association scheme for this design is then as follows: Two treatments are said to be j th associates if they differ in exactly j components. Hence we have m = ν associate classes, and the number of j th associates is ν nj = (q − 1) (j = 1, 2, . . . , ν) j Following Shah (1958), the general element of the P k matrix (k = 0, 1, . . . , ν) is given as ν − k k u+k+i −ν k pij = u ν−i −u ν−j −u u × (q − 1)ν−k−u (q − 2)k+i+j +2u−2ν

(4.55)

150 18,19,20,30,31,32

13,14,15, 25,26,27

28,32

25,26,27,29,30,31

12,24

9,10,11,21,22,23

33,34,35

36

1,25

17,29

16,28

13,14,16,25,26,28

.. .

18,19,20,22,23,24

2,3,4,10,11,12

5,6,7,9,10,11

15,27

13,15,16,25,27,28

.. .

17,21

1,9

8,12

5,6,8,9,10,12

14,26

14,15,16,26,27,28

101

2,3,4,26,27,28

6,7,8

5

14,15,16

1,2,3

4

7,11

5,7,8,9,11,12

13,25

100

13

1,2,4

3

6,10

26,27,28,10,11,12

011

.. .

1,3,4

2

5,9

010

.. .

2,3,4

001

1

000

Associate Classes

Table 4.1 Partial Association Scheme for EGD(7)-PBIB with t = 36 Treatments

4,8,16,20

5,9,29,33

13,21,25,33

20,24,32,36

19,23,31,35

18,22,30,34

17,21,29,33

110

1,2,3,5,6,7 13,14,15,17,18,19

6,7,8,10,11,12 30,31,32,34,35,36

18,19,20,22,23,24 30,31,32,34,35,36 17,19,20,21,23,24 29,31,32,33,35,36 17,18,20,21,22,24 29,31,32,33,34,36 17,18,19,21,22,23 29,30,31,33,34,35 14,15,16,22,23,24 26,27,28,34,35,36

111

151


where i, j = 0, 1, . . . , ν and the summation extends over all integer values of u such that u ≤ min(ν − i, ν − j, ν − k) and u ≥ 12 (2ν − i − j − k) k and if for a given combination i, j, k no such u value exists, then pij = 0. Specifically, from (4.55) we obtain for ν = 3 the P matrices for the cubic association scheme of Section 4.5.8.

Example 4.13 Let t = 16, q = 2, ν = 4. The treatments are 1 ≡ (1, 1, 1, 1) 5 ≡ (1, 2, 1, 1) 9 ≡ (2, 1, 1, 1) 13 ≡ (2, 2, 1, 1)

2 ≡ (1, 1, 1, 2) 6 ≡ (1, 2, 1, 2) 10 ≡ (2, 1, 1, 2) 14 ≡ (2, 2, 1, 2)

3 ≡ (1, 1, 2, 1) 7 ≡ (1, 2, 2, 1) 11 ≡ (2, 1, 2, 1) 15 ≡ (2, 2, 2, 1)

4 ≡ (1, 1, 2, 2) 8 ≡ (1, 2, 2, 2) 12 ≡ (2, 1, 2, 2) 16 ≡ (2, 2, 2, 2)

and the association scheme follows: 0th Associate 1st Associates 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

2, 1, 1, 2, 1, 2, 3, 4, 1, 2, 3, 4, 5, 6, 7, 8,

3, 4, 4, 3, 6, 5, 5, 6, 10, 9, 9, 10, 9, 10, 11, 12,

5, 6, 7, 8, 7, 8, 8, 7, 11, 12, 12, 11, 14, 13, 13, 14,

9 10 11 12 13 14 15 16 13 14 15 16 15 16 16 15

2nd Associates 4, 3, 2, 1, 2, 1, 1, 2, 2, 1, 1, 2, 1, 2, 3, 4,

6, 5, 5, 6, 3, 4, 4, 3, 3, 4, 4, 3, 6, 5, 5, 6,

7, 8, 8, 7, 8, 7, 6, 5, 5, 6, 7, 8, 7, 8, 8, 7,

10, 9, 9, 10, 9, 10, 11, 12, 12, 11, 10, 9, 10, 9, 9, 10,

11, 12, 12, 11, 14, 13, 13, 14, 14, 13, 13, 14, 11, 12, 12, 11,

13 14 15 16 15 16 16 15 15 16 16 15 16 15 14 13

3rd Associates 4th Associates 8, 7, 6, 5, 4, 3, 2, 1, 4, 3, 2, 1, 2, 1, 1, 2,

12, 11, 10, 9, 10, 9, 9, 10, 6, 5, 5, 6, 3, 4, 4, 3,

14, 13, 13, 14, 11, 12, 12, 11, 7, 8, 8, 7, 8, 7, 6, 5,

15 16 16 15 16 15 14 13 16 15 14 13 12 11 10 9

16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1

4.6.11 Right-Angular PBIB(4) Designs The right-angular association scheme was introduced by Tharthare (1963) for t = 2sq treatments. The treatments are arranged in q right angles with arms of length s, keeping the angular positions of the right angles blank. The association scheme is then as follows: Any two treatments on the same arm are first associates; any two treatments on different arms of the same right angle are second associates;

152


any two treatments on the same (i.e., parallel) arm but in different right angles are third associates; any two treatments are fourth associates otherwise. It follows then that n1 = s − 1

n2 = s

n3 = (q − 1)s = n4

and 

0 1 1 s − 2   0 P 1 = 0  0 0 0 0

 0 0 0 0 0 0    s 0 0   0 s(q − 1) 0  0 0 s(q − 1)

 0 0 1 0 0 0 0 s−1 0 0      0 0 0  P 2 = 1 s − 1   0 0 0 0 s(q − 1) 0 0 0 s(q − 1) 0 



 0 1 0 0 s −1 0    0 0 s   0 s(q − 2) 0  s 0 s(q − 1)



 0 0 1 0 0 s −1    0 s 0   s 0 s(q − 2) 0 s(q − 2) 0

0 0 0 0   0 P 3 = 0  1 s − 1 0 0 0 0 0 0   0 P 4 = 0  0 0 1 s −1

Example 4.14 Let t = 12, q = 3, s = 2. Treatments: 1

5

9

2

6

10

3

4

7

8

11

12

153


The association scheme follows: 0th Associate

1st Associates

2nd Associates

3rd Associates

4th Associates

1 2 3 4 5 6 7 8 9 10 11 12

2 1 4 3 6 5 8 7 10 9 12 11

3, 4 3, 4 1, 2 1, 2 7, 8 7, 8 5, 6 5, 6 11, 12 11, 12 9, 10 9, 10

5, 6, 9, 10 5, 6, 9, 10 7, 8, 11, 12 7, 8, 11, 12 1, 2, 9, 10 1, 2, 9, 10 3, 4, 11, 12 3, 4, 11, 12 1, 2, 5, 6 1, 2, 5, 6 3, 4, 7, 8 3, 4, 7, 8

7, 8, 11, 12 7, 8, 11, 12 5, 6, 9, 10 5, 6, 9, 10 3, 4, 11, 12 3, 4, 11, 12 1, 2, 9, 10 1, 2, 9, 10 3, 4, 7, 8 3, 4, 7, 8 1, 2, 5, 6 1, 2, 5, 6

Tharthare (1965) extended this association scheme to what he called the generalized right-angular association scheme for t = pqs treatments with p > 2. To describe the association scheme succinctly, it is useful to denote the treatment by triplets (x1 , x2 , x3 ) with x1 = 1, 2, . . . , q; x2 = 1, 2, . . . , p; x3 = 1, 2, . . . , s. Then, two treatments (x1 , x2 , x3 ) and (y1 , y2 , y3 ) are said to be 1st Associates: if x1 = y1 , x2 = y2 , x3 = y3 2nd Associates: if x1 = y1 , x2 = y2 3rd Associates: if x1 = y1 , x2 = y2 4th Associates: otherwise This, obviously, leads = (q − 1)(p − 1)s.

to

n1 = s − 1, n2 = (p − 1)s, n3 = (q − 1)s, n4

4.6.12 Cyclic PBIB Designs Cyclic PBIB designs represent a rather large and flexible class of incomplete block designs. They were introduced by Kempthorne (1953) and Zoellner and Kempthorne (1954) for blocks of size k = 2. Further developments are due to for example, David (1963, 1965), David and Wolock (1965), John (1966, 1969), and Wolock (1964). Let us denote the treatments by 0, 1, 2, . . . , t − 1. The association scheme can then be defined as follows, where we need to distinguish between t even and t odd: t even: For a fixed treatment θ , the uth associates are (θ + u, θ − u)mod t(u = 1, 2, . . . , t/2 − 1) and the t/2th associate is θ + t/2 (mod t). Thus we have m = t/2 associate classes with nu = 2 (u = 1, 2, . . . , t/2 − 1) and nt/2 = 1.

154


t odd : The uth associates of θ are (θ + u, θ − u)mod t for u = 1, 2, . . . , (t − 1)/2. Thus we have m = (t − 1)/2 associate classes with nu = 2 [u = 1, 2, . . . , (t − 1)/2]. Example 4.15 Let t = 6. 0th Associate

1st Associates

0 1 2 3 4 5

1, 2, 3, 4, 5, 0,

2nd Associates

5 0 1 2 3 4

2, 3, 4, 5, 0, 1,

4 5 0 1 2 3

3rd Associates 3 4 5 0 1 2

We draw the reader’s attention to the cyclic development of the treatments in a given associate class, hence the name of the association scheme. The P matrices are as follows:   0 1 0 0 1 0 1 0  P1 =  0 1 0 1 0 0 1 0 

0 0 P2 =  1 0 

0 0 P3 =  0 1

0 1 0 1

1 0 1 0

 0 1  0 0

0 0 2 0

0 2 0 0

 1 0  0 0

4.6.13 Some Remarks We conclude this section with the following remarks: 1. As we have mentioned earlier, the list of association schemes given here is not exhaustive. Other association schemes with m > 2 associate classes were given by for example, Roy (1953–1954), Raghavarao (1960), Ogasawara (1965), Yamamoto, Fuji, and Hamada (1965), and Adhikary (1966, 1967). In most cases they represent generalizations of existing association schemes.

ESTIMATION OF ρ FOR PBIB(2) DESIGNS

155

2. We have chosen our list of association schemes because they lead to a rather large class of existing and practical PBIB designs [in particular the PBIB(2) designs and the cyclic PBIB designs are well documented (see Chapter 5)] and/or are particularly useful in the construction of systems of confounding for symmetrical and asymmetrical factorial designs (see Chapters 11 and 12). 3. We emphasize that an association scheme does not constitute a PBIB design, nor does it lead automatically to a PBIB design. Methods of constructing designs with a certain association scheme is the subject of Chapter 5. It will turn out that certain association schemes are connected with certain construction methods. 4.7 ESTIMATION OF ρ FOR PBIB(2) DESIGNS Recall from Sections 1.10 and 2.5 that the general problem with utilizing the combined intra- and interblock information is that of finding an estimator for ρ, ρ , such that var[t ( ρ )] ≤ var(t)

(4.56)

where t is the intrablock estimator for the estimable function c τ and t ( ρ ) is the combined estimator for the same function using ρ as the estimate of ρ. Shah (1964) has derived as estimator for ρ for a certain class of designs such that (4.56) holds. We shall state his result without proof and then apply it to the various types of PBIB(2) designs as in Section 4.6. 4.7.1 Shah Estimator Let D1 be the class of proper, equireplicate, binary incomplete block designs for which the concordance matrix NN has only one nonzero eigenvalue (other than rk). Let θ denote this eigenvalue with multiplicity α = rank(N N ) − 1. Further, define Z = c (1 + c){SS(X τ | I) − (2T − r τ ) τ}

(4.57)

where c = (rk/θ ) − 1 and all the other terms are as defined in earlier sections. We now state the following theorem. Theorem 4.1 Consider an incomplete block design belonging to the class D1 . When  Z rkα Z θ   −1 if >  MSE θ ρ = rk − θ αMSE   1 otherwise

156


with Z as given in (4.57), is used in the combined analysis, then for any treatment contrast c τ , var[t ( ρ )] < var(t) for all values of ρ provided that (α − 4)(n − t − b − 1) ≥ 8

(4.58)

4.7.2 Application to PBIB(2) Designs We have already used the result of Theorem 4.1 in connection with BIB designs (see Section 2.5) that clearly belong to D1 . We have mentioned earlier that among all PBIB designs the PBIB(2) designs are quite numerous and most useful from a practical point of view. As we have seen in Section 4.4.3, they have two distinct eigenvalues other than rk and expressions for these and their multiplicities are given in (4.31) and (4.32), respectively. Using these results we shall now investigate which PBIB(2) designs belong to the class D1 and give for them the corresponding condition (4.58). 1. Group-Divisible Designs Using the explicit expressions for P 1 and P 2 as given in Section 4.6.1, we find that θ1 = r − λ1

α1 = t1 (t2 − 1)

θ2 = r − λ1 + t2 (λ1 − λ2 )

α2 = t1 − 1

In particular, we have, following Bose and Connor (1952), for a. Singular GD designs: r − λ1 = 0. Hence the designs belong to D1 , with θ = θ2 and α = α2 . Condition (4.58) then becomes (t1 − 5)(n − b − t − 1) ≥ 8. b. Semiregular GD designs: r − λ1 + t2 (λ1 − λ2 ) = 0. Hence these designs also belong to D1 , with θ = θ1 and α = α1 . Condition (4.58) can, therefore, be written as [t1 (t2 − 1) − 4](n − b − t − 1) ≥ 8. c. Regular GD designs: They do not belong to D1 since θ1 > 0 and θ2 > 0. 2. Triangular Designs Using the results of Section 4.6.2 we obtain θ1 = r + (q − 4)λ1 − (q − 3)λ2 θ2 = r − 2λ1 + λ2

α1 = q − 1

α2 = q(q − 3)/2

They belong to D1 if either one of these eigenvalues is zero, which is possible.

ESTIMATION OF ρ FOR PBIB(2) DESIGNS

157

3. Latin Square Type Designs 4.6.3) we have

For the Li -association scheme (see Section

θ1 = r − (i − q)(λ1 − λ2 ) − λ2 θ2 = r − i(λ1 − λ2 ) − λ2

α1 = i(q − 1)

α2 = (q − 1)(q − i + 1)

they belong to D1 if either θ1 or θ2 is zero, which is possible. 4. Cyclic Designs Neither the expressions for the eigenvalues or their multiplicities reduce to simple expressions. To check whether a given design belongs to D1 , we need to work out (4.31) and check whether either θ1 or θ2 equals zero.

CHAPTER 5

Construction of Partially Balanced Incomplete Block Designs

In the previous chapter we discussed several types of association schemes for PBIB designs with two or more associate classes. There exists a very large number of PBIB designs with these, and other, association schemes. But as we have pointed out earlier, the association schemes themselves do not constitute or generate the actual plan, that is, the assignment of treatment to blocks. To obtain such plans, different methods of construction have been developed and used. It is impossible to mention all of them as the literature in this area is immense. Rather, we shall discuss in some detail only a few methods with emphasis mainly on three types of PBIB designs, namely PBIB(2) designs, in particular, group-divisible PBIB(2) designs, cyclic PBIB designs, and EGD-PBIB designs. The first class is important because it constitutes the largest class of PBIB designs. The second class also represents a rather large class, and the designs are easy to construct and are widely applicable. The last class is important with respect to the construction of systems of confounding for asymmetrical factorial experiments (see Chapter 12). 5.1 GROUP-DIVISIBLE PBIB(2) DESIGNS We shall give a brief discussion of the basic methods for constructing groupdivisible designs. These methods are (i) duals of BIB designs, (ii) method of differences, (iii) finite geometries, and (iv) orthogonal arrays, as developed mainly by Bose, Shrikhande, and Bhattacharya (1953). 5.1.1 Duals of BIB Designs Let N be the incidence matrix of a design D with parameters t, b, k, r. Then N denotes the incidence matrix of a design D , with parameters t , b , k , r , which Design and Analysis of Experiments. Volume 2: Advanced Experimental Design By Klaus Hinkelmann and Oscar Kempthorne ISBN 0-471-55177-5 Copyright  2005 John Wiley & Sons, Inc.

158

159

GROUP-DIVISIBLE PBIB(2) DESIGNS

has been obtained from D by interchanging treatments and blocks. This implies that D has parameters t = b, b = t, k = r, r = k. The design D is referred to as the dual of the design D (and, of course, vice versa). If the design D has a certain structure, such as being a BIB or PBIB design, then the dual design D very often also has a particular recognizable structure. It is this relationship between D and D that we shall now utilize to construct certain GD-PBIB(2) designs from existing BIB designs. Consider the following theorem. Theorem 5.1 Let D be a BIB design with parameters t = s 2 , b = s(s + 1), k = s, r = s + 1, λ = 1, where s is prime or prime power (see Sections 3.34 and 18.7). Then the dual D is a GD-PBIB(2) design with parameters t = (s + 1)s, t1 = s + 1, t2 = s, b = s 2 , k = s + 1, r = s, λ1 = 0, λ2 = 1. Proof As can be seen from the construction in Section 3.7, D is a resolvable (α = 1) BIB design, consisting of s + 1 replication groups with s blocks each and containing each treatment once. Denote the blocks by Bj where j denotes the replication group (j = 1, 2, . . . , s + 1) and denotes the block within the j th replication group ( = 1, 2, . . . , s). Now consider a treatment θ , say, and suppose it occurs in blocks B11 , B22 , . . . , Bs+1,s+1 . Then in D block θ contains the treatments 11 , 22 , . . . , s + 1s+1 . Further, let θ1 , θ2 , . . . , θs−1 be the treatments that occur together with θ in B11 in D. Since {θ, θ1 , θ2 , . . . , θs−1 } appear all in different blocks in the remaining s replication groups (since λ = 1), it follows then that in D treatment 11 occurs together exactly once with all j (j = 2, . . . , s + 1; = 1, 2, . . . , s), but not at all with the s − 1 treatments 1( = 1 ). This implies the association scheme 11 21

12 22

... ...

1s 2s

s+1 1

s+1 2

...

s+1 s

where treatments in the same row are 1st associates and 2nd associates otherwise, which is, in fact, the GD association scheme, and since the GD association is unique (Shrikhande, 1952), the resulting design is the GD-PBIB(2) design. Example 5.1 Consider the BIB design D with t = 32 , b = 12, k = 3, r = 4, λ = 1, given by the plan [denoting the treatments by (x1 , x2 ), x1 , x2 = 0, 1, 2]: Block

Treatments

11 12 13

(0, 0) (1, 0) (2, 0)

(0, 1) (1, 1) (2, 1)

(0, 2) (1, 2) (2, 2)

21 22 23

(0, 0) (0, 1) (0, 2)

(1, 0) (1, 1) (1, 2)

(2, 0) (2, 1) (2, 2)

160

CONSTRUCTION OF PARTIALLY BALANCED INCOMPLETE BLOCK DESIGNS

Block

Treatments

31 32 33

(0, 0) (1, 0) (2, 0)

(1, 2) (0, 1) (0, 2)

(2, 1) (2, 2) (1, 1)

41 42 43

(0, 0) (1, 0) (2, 0)

(1, 1) (0, 2) (0, 1)

(2, 2) (2, 1) (1, 2)

The dual D is then given by Block

Treatments

(0, 0) (0, 1) (0, 2)

11, 21, 31, 41 11, 22, 32, 43 11, 23, 33, 42

(1, 0) (1, 1) (1, 2)

12, 21, 32, 42 12, 22, 33, 41 12, 23, 31, 43

(2, 0) (2, 1) (2, 2)

13, 21, 33, 43 13, 22, 31, 42 13, 23, 32, 41

The reader can verify easily that this is indeed the plan for a GD-PBIB(2) design with the parameters as indicated. It is isomorphic to plan SR41 given by Clatworthy (1973). 5.1.2 Method of Differences We denote and write the t = t1 t2 treatments now as follows: 01 11 (t1 − 1)1

02 12 (t1 − 1)2

... ... ...

0t2 1t2 (t1 − 1)t2

(5.1)

that is, the (i + 1)th group in the association scheme for the GD design consists of the treatments (i1 , i2 , . . . , it2 ) with i = 0, 1, . . . , t1 − 1, these treatments being 1st associates of each other. The treatments in the th column are referred to as treatments of class . The method of differences as described for BIB designs (Section 3.2) can now be extended for purposes of constructing GD-PBIB(2) designs as stated in the following theorem.

161


Theorem 5.2 Let it be possible to find s initial blocks B10 , B20 , . . . , Bs0 such that i. Each block contains k treatments. ii. Treatments of class ( = 1, 2, . . . , t2 ) are represented r times among the s blocks. iii. Among the differences of type (, ) arising from the s blocks, each nonzero residue mod t1 occurs the same number of times, λ2 , say, for all = 1, 2, . . . , t2 . iv. Among the differences of type (, ) arising from the s blocks, each nonzero residue mod t1 occurs λ2 times and 0 occurs λ1 times, say, for all (, )(, = 1, 2, . . . , t2 , = ). Then, developing the initial blocks B10 , B20 , . . . , Bs0 cyclically, mod t1 yields a GD-PBIB(2) design with parameters t = t1 t2 , b = t1 s, k, r = ks/t2 , λ1 , λ2 . The proof is quite obvious as it is patterned after that of Theorem 3.1. By looking at array (5.1), it is clear that through the cyclic development of the initial blocks, any two treatments in the same row occur λ1 times together in the same block, and any two treatments not in the same row occur λ2 times together in the same block. This then satisfies the association scheme for the GD-PBIB(2) design. We shall illustrate this with the following example. Example 5.2 Let t = 14, t1 = 7, t2 = 2; that is, array (5.1) is 01 11 21 31 41 51 61

02 12 22 32 42 52 62

Let the initial blocks of size 4 be B10 = (01 , 02 , 11 , 12 ) B20 = (01 , 02 , 21 , 22 ) B30 = (01 , 02 , 31 , 32 ) To verify conditions 2–4 of Theorem 5.2 we see that treatments of class 1 and 2 occur r = 6 times. Further, the differences of type (1, 1) are B10 :

01 − 11 ≡ 6

11 − 01 ≡ 1

B20 :

01 − 21 ≡ 5

21 − 01 ≡ 2

B30 :

01 − 31 ≡ 4

31 − 01 ≡ 3

162


and the same holds for differences of type (2, 2); similarly, differences of type (1, 2) are B10 : 01 − 02 ≡ 0

01 − 12 ≡ 6

11 − 02 ≡ 1

11 − 12 ≡ 0

B20 : 01 − 02 ≡ 0

01 − 22 ≡ 5

21 − 02 ≡ 2

21 − 22 ≡ 0

B30 : 01 − 02 ≡ 0

01 − 32 ≡ 4

31 − 02 ≡ 3

31 − 32 ≡ 0

and the same is true for differences of type (2, 1). Hence λ1 = 6, λ2 = 1. The plan for the GD-PBIB(2) design with b = 21 blocks is then (01 , 02 , 11 , 12 ) (11 , 12 , 21 , 22 ) (21 , 22 , 31 , 32 ) (31 , 32 , 41 , 42 ) (41 , 42 , 51 , 52 ) (51 , 52 , 61 , 62 ) (61 , 62 , 01 , 02 )

(01 , 02 , 21 , 22 ) (11 , 12 , 31 , 32 ) (21 , 22 , 41 , 42 ) (31 , 32 , 51 , 52 ) (41 , 42 , 61 , 62 ) (51 , 52 , 01 , 02 ) (61 , 62 , 11 , 12 )

(01 , 02 , 31 , 32 ) (11 , 12 , 41 , 42 ) (21 , 22 , 51 , 52 ) (31 , 32 , 61 , 62 ) (41 , 42 , 01 , 02 ) (51 , 52 , 11 , 12 ) (61 , 62 , 21 , 22 )

which is the same as plan S13 of Clatworthy (1973). We note that this PBIB design is, of course, resolvable and that it lends itself to two-way elimination of heterogeneity in that each treatment occurs exactly three times in the first two and the last two positions of the 21 blocks. 5.1.3 Finite Geometries We shall first consider a method of using a projective geometry, PG(K, p n ), to construct GD-PBIB(2) designs. The result can then be applied similarly to a Euclidean geometry, EG(K, p n ). (See Appendix B for details about projective and Euclidean geometries). The basic idea in both cases is to omit one point from a finite geometry and all the M-spaces containing that point. The remaining M-spaces are then taken as the blocks of a PBIB design. More specifically we have the following theorem. Theorem 5.3 Omitting one point from a PG(K, p n ) and all the M-spaces containing that point, gives rise to a GD-PBIB(2) design, if one identifies the remaining points with the treatments and the remaining M-spaces with the blocks. Any two treatments are 1st associates if the line joining them goes through the omitted point, they are 2nd associates otherwise. Proof The PG(K, p n ) has 1 + p n + p 2n + · · · + p Kn points. Omitting one point leads to t = p n + p 2n + · · · + p Kn = 1 + p n + · · · + p (K−1)n p n ≡ t1 t2

(5.2)

163


treatments. In the PG(K, p n ) there are ψ(M, K, p n ) M-spaces, of which ϕ(0, M, K, p n ) contain the omitted point. Hence there are b = ψ(M, K, p n ) − ϕ(0, M, K, p n )

(5.3)

M-spaces left that constitute the blocks, each M-space (block) containing k = ψ(0, M, p n ) = 1 + p n + p 2n + · · · + p Mn

(5.4)

points (treatments). Since each point is contained in ϕ(0, M, K, p n ) M-spaces and since each line joining a given point and the omitted point is contained in ϕ(1, M, K, p n ) M-spaces, it follows that each retained point (treatment) occurs in r = ϕ(0, M, K, p n ) − ϕ(1, M, K, p n )

(5.5)

retained M-spaces (blocks). Further n1 = ψ(0, 1, p n ) − 2 = p n − 1 = t2 − 1 n2 = t − n1 − 1 = p n + p 2n + · · · + p (K−1)n p n = (t1 − 1)t2 and

(5.6)

(5.7)

λ1 = 0, λ2 = ϕ(1, M, K, p n )

(5.8)

Thus, we are obviously led to the association scheme for the GD-PBIB(2) design. Because of the uniqueness of the association scheme, the design constructed in this manner is a GD-PBIB design. As an illustration we consider the following example. Example 5.3 Let p = 3, K = 2, n = 1. The points of the PG(2, 2) are given by triplets (u1 , u2 , u3 ) with u1 , u2 , u3 = 0, 1 except (u1 u2 u3 ) = (000). Suppose we omit the point (100), then the remaining points are: 010, 110, 001, 101, 011, 111, that is, t = 6. The lines (i.e., M = 1) not passing through 100 constitute the blocks of the design and are given by 001

010

011

110

001

111

010

101

111

110

101

011

(5.9)

that is, b = 4, k = 3, r = 2. The association scheme is obtained by determining for each point which other points lie on the line going through it and the omitted

164


point. For example, the only point on the line going through (001) and (100) is obtained from (µ0 + ν1, µ0 + ν0, µ1 + ν0) with µ = ν = 1, that is, the point (101). Doing this for all six points leads to the following association scheme: 0th Associate

1st Associates

010 110 001 101 011 111

110 010 101 001 111 011

2nd Associates 001, 001, 010, 010, 010, 010,

101, 101, 110, 110, 110, 110,

011, 011, 011, 011, 001, 001,

111 111 111 111 101 101

This is, of course, the GD-PBIB association scheme if we write the treatments in the following array: 010 001 011

110 101 111

Inspection of the plan (5.9) shows that λ1 = 0, λ2 = 1. With suitable labeling this is plan SR18 of Clatworthy (1973). 5.1.4 Orthogonal Arrays Orthogonal arrays (for a description see Appendix C) play an important role in the construction of experimental designs (see also Chapters 13 and 14). Of particular interest are orthogonal arrays of strength 2. This is the type we shall employ here to construct GD-PBIB designs. More specifically we shall employ orthogonal arrays OA[N, K, p, 2; λ], where p is prime. Their relationship to the GD-PBIB(2) design is as follows: Replace any integer x appearing in the ith row of the array by the treatment (i − 1)p + x. The ith row contains then the treatments (i − 1)p, (i − 1)p + 1, . . . , (i − 1)p + p − 1 each occurring r = N/p times since each symbol occurs equally often in each row. The total number of treatments is t = Kp and we have t1 = K, t2 = p. The columns of this derived scheme form the blocks of the PBIB design, that is, there are b = N blocks, each of size k = K. Treatments in the same row are 1st associates, that is, each treatment has n1 = p − 1 1st associates and since these treatments do not appear in any other row, it follows that λ1 = 0. Treatments in different rows are 2nd associates, giving n2 = (K − 1)p. Since we have an OA[b, k, p, 2; λ], that is, each possible 2 × 1 column vector appearing λ times, we have λ2 = λ. This is, of course, the association scheme of a GD-PBIB(2) design with the parameters as stated above.

165

CONSTRUCTION OF OTHER PBIB(2) DESIGNS

We illustrate this with the following example. Example 5.4 Consider the  0 0 A= 0 0

following OA[8, 4, 2, 2, 2]:  1 0 1 0 1 0 1 0 1 1 0 0 1 1  0 0 0 1 1 1 1 1 1 0 1 0 0 1

This leads to the derived GD-PBIB(2) design (with columns representing blocks): 1

2

3

Block 4 5

6

7

8

0 2 4 6

1 2 4 7

0 3 4 7

1 3 4 6

1 2 5 6

0 3 5 6

1 3 5 7

0 2 5 7

with t = 8, b = 8, k = 4, r = 4, λ1 = 0, λ2 = 2. This design is isomorphic to plan SR36 of Clatworthy (1973). For further specialized methods of constructing GD-PBIB(2) designs the reader is referred to Raghavarao (1971). An essentially complete list of existing plans for GD-PBIB(2) designs is given by Clatworthy (1973) together with references and methods concerning their construction. 5.2 CONSTRUCTION OF OTHER PBIB(2) DESIGNS We shall mention here only a few methods but otherwise refer the reader to the pertinent literature. 5.2.1 Triangular PBIB(2) Designs Recall that the number of treatments is t = q(q − 1)/2 and that for the association scheme the treatments are arranged in a triangular array and its mirror image (see Section 4.6.2). One method of constructing triangular designs can then be described as follows. If we let each row of the association scheme be a block, then the resulting design is a triangular PBIB design with parameters: t = q(q − 1)/2

b=q

k =q −1

r=2

λ1 = 1

Example 5.5 Let t = 6 = 4 × 32 . The association scheme is * 1 2 3

1 * 4 5

2 4 * 6

3 5 6 *

λ2 = 0

166


The design then is given by 1: 2: 3: 4:

Block

1 1 2 3

2 4 4 5

3 5 6 6

that is, b = 4, k = 3, r = 2, λ1 = 1, λ2 = 0.

Other methods given by Shrikhande (1960, 1965) and Chang, Liu, and Liu (1965) are based on the existence of certain BIB designs and considering, for example, the dual of the BIB design as omitting certain blocks from the BIB design. Still other methods are given by Clatworthy (1955, 1956), Masuyama (1965), and Ray-Chaudhuri (1965). A listing of practically useful designs is given by Clatworthy (1973). 5.2.2 Latin Square PBIB(2) Designs Similar to the procedure described in the previous section, the following method of constructing L2 -PBIB designs is closely connected to the square treatment array for the L2 -association scheme. We have t = q 2 , and suppose q is prime or prime power. We know that there exist then q − 1 mutually orthogonal Latin squares (MOLS) of order q (see Section I.10.6.2). The languages for these q − 1 MOLS are then superimposed on the treatment array, that is, each treatment has associated with it q − 1 letters, one from each of the q − 1 MOLS. Collecting the treatments that have the same letter from each of the q − 1 languages and calling those sets blocks, leads to an L2 -PBIB design with the following parameters: t = q2

b = q(q − 1)

k=q

r =q−1

λ1 = 0

λ2 = 1

Example 5.6 Let t = 16, q = 4. The 4 × 4 association scheme with the superimposed three MOLS is as follows: 1 Aα I 5 B γ IV 9 C δ II 13 D β III

2 B β II 6 A δ III 10 Dγ I 14 C α IV

3 C γ III 7 D α II 11 A β IV 15 Bδ I

4 D δ IV 8 Cβ I 12 B α III 16 A γ II

167

CYCLIC PBIB DESIGNS

The blocks are then Block 1(A) 2(B) 3(C) 4(D) 5(α) 6(β) 7(γ ) 8(δ) 9(I) 10(II) 11(III) 12(IV)

Treatments 1, 2, 3, 4, 1, 2, 3, 4, 1, 2, 3, 4,

6, 5, 8, 7, 7, 8, 5, 6, 8, 7, 6, 5,

11, 12, 9, 10, 12, 11, 10, 9, 10, 9, 12, 11,

16 15 14 13 14 13 16 15 15 16 13 14

It is, of course, obvious from the method of construction that this PBIB design is a resolvable design with the blocks from each language forming a complete replicate. Other methods of construction, some based on the existence of certain BIB designs, have been given by for example, Bose, Clatworthy, and Shrikhande (1954), Chang and Liu (1964), Clatworthy (1955, 1956, 1967), and, as lattice designs, by Yates (1936b) (see Section 18.7). A list of these designs can be found in Clatworthy (1973). 5.3 CYCLIC PBIB DESIGNS We shall now turn to the rather large class of cyclic PBIB designs we introduced in Section 4.6.12. These designs are quite useful since (1) they are easy to construct, namely as the name suggests through cyclic development of initial blocks, (2) they exist for various combinations of design parameters, and (3) they are easy to analyze, that is, their analytical structure is easy to derive. 5.3.1 Construction of Cyclic Designs As mentioned earlier, the construction of cyclic designs is based on the cyclic development of a set of initial blocks. We can distinguish basically between four types of cyclic designs according to whether the number of blocks b is (1) b = t, (2) b = st, (3) b = t/d, or (4) b = t (s + 1/d), where d is a divisor of t. It is then obvious that For (1) we need one initial block of size k; For (2) we need s distinct (nonisomorphic) initial blocks of size k;

168


For (3) we need one initial block of size k such that, if cyclically developed, block b + 1 = t/d + 1 is the same as the initial block and each treatment is replicated r times. This means that if the initial block contains treatment i, it also must contain treatments i + b, i + 2b, . . . , i + (d − 1)b. This implies that d is also a divisor of k, say k/d = k . Then any k treatments i1 , i2 , . . . , ik can be chosen and the remaining treatments in the initial block are determined; For (4) we combine initial blocks from type (1) if s = 1 or (2) if s > 1 with an initial block from type (3). The choice of initial blocks for (1), (2), and hence (4) above is quite arbitrary since any choice will lead to an appropriate design. There is, however, one other consideration that plays an important role in choosing initial blocks and that is the efficiency E of the resulting design. Efficiency here is defined in terms of the average variance of all treatment comparisons (see Section 1.10), and high efficiency is closely related to a small number of associate classes. Whereas the number of associate classes for a cyclic design is, in general, m = t/2 for t even and m = (t − 1)/2 for t odd, this number can sometimes be reduced by a proper choice of initial blocks, in some cases even to m = 1, that is, BIB designs, or m = 2, that is, PBIB(2) designs. For example, consider t = 6, k = 3, r = 3, b = 6: If the initial block is (1, 2, 3), then the plan is 1 2 3 4 5 6

2 3 4 5 6 1

3 4 5 6 1 2

and by inspection we can establish the following association scheme: 0th Associate 1 2 3 4 5 6 with λ1  0 1 P1 =  0 0

1st Associates 2, 3, 4, 5, 6, 1,

2nd Associates

6 1 2 3 4 5

= 2, λ2 = 1, λ3 = 0, and   1 0 0 0 0 0 1 0  P2 =  1 1 0 1 0 1 0 0

3, 4, 5, 6, 1, 2,

0 1 0 1

1 0 1 0

 0 1  0 0

3rd Associates

5 6 1 2 3 4

4 5 6 1 2 3



0 0 P3 =  0 1

0 0 2 0

0 2 0 0

 1 0  0 0

169

CYCLIC PBIB DESIGNS

Hence we have a PBIB(3) design. If, on the other hand, the initial block is (1, 2, 4), then the resulting plan is 1 2 3 4 5 6

2 3 4 5 6 1

4 5 6 1 2 3

and, by inspection, we can speculate on the association scheme as being 0th Associate 1 2 3 4 5 6

1st Associates 2, 3, 4, 5, 6, 1,

with λ1 = 1, λ2 = 2, and   0 1 0 P 1 = 1 2 1  0 1 0

3, 4, 5, 6, 1, 2,

5, 6, 1, 2, 3, 4,

2nd Associates

6 1 2 3 4 5

4 5 6 1 2 3

 0 0 1 P 2 = 0 4 0 1 0 0 

Hence we have in fact a PBIB(2) design. For the first design we obtain E = 0.743 [see (5.14)], whereas for the second design we find E = 0.784. Hence the second design is slightly preferable from an efficiency point of view, and this design is therefore listed by John, Wolock, and David (1972). One can convince oneself that no better cyclic design for this combination of parameters exists. The association scheme for the second design is, indeed, the association scheme for a cyclic PBIB(2) design as discussed in Section 4.6.4, except that we have to relabel the treatments as 0, 1, . . . , 5. We then have d1 = 1, d2 = 2, d3 = 4, d4 = 5, e1 = 3, α = 2, β = 4. An extensive list of initial blocks for cyclic designs for various parameters with 6 ≤ t ≤ 30, k ≤ 10, r ≤ 10, and fractional cyclic designs for 10 ≤ t ≤ 60, 3 ≤ k ≤ 10 is given by John, Wolock, and David (1972). A special group in this collection of plans are those with k = 2, which are often referred to as paired comparison designs. 5.3.2 Analysis of Cyclic Designs The general methods of analyzing incomplete block designs (see Chapter 1) apply, of course, also to the cyclic designs just described, or even more specifically, the methods of analyzing BIB designs (where appropriate) or PBIB(m)

170


designs can be used. However, as pointed out earlier, for many cyclic designs the number of associate classes, m, is rather large and hence the system of equations (4.34) becomes rather large. Instead, another method of solving the RNE C τˆ = Q for cyclic designs can be used. This method is interesting in itself as it makes use of the particular construction process that leads to a circulant matrix C and hence to a circulant matrix C − , the elements of which can be written out explicitly (Kempthorne, 1953). Let us denote the first row of C by (c1 , c2 , . . . , ct ). The second row is then obtained by shifting each element of the first row one position to the right (in a circulant manner), and so forth so that C is determined entirely by its first row. The same is true for C − , so all we need to know in order to solve the RNE and to obtain expressions for variances of estimable functions λ τ , is the first row of C − , which we denote by c1 , c2 , . . . , ct . Let dj (j = 1, 2, . . . , t) denote the eigenvalues of C where (Kempthorne, 1953) dj =

t

c ( − 1)(j − 1)θ

(5.10)

=1

and θ = 2π/t. Note that d1 = 0; hence d2 , d3 , . . . , dt are the nonzero eigenvalues of C. Then, as shown by Kempthorne (1953), the elements ci (i = 1, 2, . . . , t) are given as 1 1 t dj t

c1 =

j =2

t 1 cos(j − 1)(i − 1)θ ci = t dj

(5.11) (i = 2, 3, . . . , t)

j =2

Expressions (5.10) and (5.11) can actually be simplified somewhat. Because of the construction of the designs and the resulting association scheme, we have for the λ1i in N N λ12 = λ1t , λ13 = λ1,t−1 , λ1,t/2 = λ1,t/2+2 λ14 = λ1,t−2 , . . . , λ1,(t+1)/2 = λ1,(t+3)/2 and hence we have for C k−1 k −λ12 c2 = ct = k

c1 = r

for t even for t odd

171

CYCLIC PBIB DESIGNS

−λ13 k −λ14 c4 = ct−2 = k −λ1,(t+1)/2 c(t+1)/2 = c(t+3)/2 = k c3 = ct−1 =

or c(t+2)/2 =

−λ1,(t+2)/2 k

for t odd for t even

As a consequence, also the maximum number of different ci values is (t + 1)/2 for t odd or (t + 2)/2 for t even. With the ci values given in (5.11) we can find the variances of treatment comparisons in general via comparisons with treatment 1 as follows: τi ) = var( τ1 − τi −i+1 ) var( τi −

= 2(c1 − ci −i+1 )σe2

for i > i

(5.12)

This is, of course, a consequence of the circulant form of C − . The average variance of all such treatment comparisons then is av. var =

1 var( τi − τi ) t (t − 1) ii i=i

=

1 var( τ1 − τi ) t −1 i=1

and using (5.12), av. var =

2 1 (c − ci )σe2 t −1 i=1

t 2 1 i tc − = c σe2 t −1

i=1

2t 1 2 = c σe t −1

(5.13)

since ti=1 ci = 0 as can be verified directly by using (5.11). From (5.12) it follows then that the efficiency factor for the treatment comparison between the ith and the i th treatment is Eii =

2σe2 /r 1 = −i+1 1 i 2 1 2(c − c )σe r(c − ci −i+1 )

172


and from (5.13) that the overall efficiency factor is E=

t −1 2σe2 /r = 1 2 2tc σe /(t − 1) rtc1

(5.14)

5.4 KRONECKER PRODUCT DESIGNS In previous sections we have used the fact that incomplete block designs can sometimes be generated from some other existing incomplete block design, for example, the residual design of a BIB design is also a BIB design, the duals of certain BIB designs are PBIB(2) designs. In this section we shall use another widely applicable and useful technique to generate new incomplete block designs from existing designs. This method and hence the resulting designs are referred to as Kronecker product designs. These designs were first introduced by Vartak (1955) and later put on a more formal basis by Surendran (1968). 5.4.1 Definition of Kronecker Product Designs Suppose we have two PBIB designs, D1 and D2 , with m1 and m2 associate classes, respectively [in what follows we consider a BIB design as a PBIB(1) design] defined by their respective incidence matrices N 1 and N 2 with the following parameters: D1 : t1 , r1 , k1 , b1 , nu , λu D2 : t2 , r2 , k2 , b2 , n∗f , λ∗f

k P k = (puv )

(u, v, k = 0, 1, . . . , m1 )

P ∗f = (qgh )

(f, g, h = 0, 1, . . . , m2 )

f

(5.15)

Consider now the Kronecker product N = N1 × N2

(5.16)

where N is obtained by multiplying each element in N 1 by the matrix N 2 . It is immediately obvious that N is an incidence matrix of dimension t × b with t = t1 t2 and b = b1 b2 . Furthermore, each row of N contains r = r1 r2 unity elements and each column contains k = k1 k2 unity elements. Hence the design D, defined by its incidence matrix (5.16), is an incomplete block design with t treatments, b blocks, r replicates per treatment, and k units per block. It remains for us to show that this Kronecker product design is indeed a PBIB design and to establish its association scheme and the remaining parameters. 5.4.2 Properties of Kronecker Product Designs Let us denote the treatments of D1 by the vector θ = (θ1 , θ2 , . . . , θt1 ) and the treatments of D2 by the vector φ = (φ1 , φ2 , . . . , φt2 ). The treatments for design D can then be defined by what Kurkjian and Zelen (1962, 1963) have called the symbolic direct product between θ and φ.

173

KRONECKER PRODUCT DESIGNS

Definition 5.1 Let θ = (θ1 , θ2 , . . . , θq ) and φ = (φ1 , φ2 , . . . , φs ) be two arrays of size q and s, respectively. Then a new array x of size q · s is created by the symbolic direct product (SDP) of θ and φ as follows: 

 x11 x12     ..   .     x =θ ⊗φ = x1s  x21     ..   .  xqs

(5.17)

It is convenient to enumerate the rows of the matrix N in (5.16) in the same way as the array in (5.17). We shall now prove the following theorem. Theorem 5.4 Let D1 and D2 be two PBIB designs with parameters as given in (5.15). Then the Kronecker product design D as defined by its incidence matrix N = N 1 × N 2 is a PBIB design with m = (m1 + 1)(m2 + 1) − 1 associate classes and the following association scheme: If treatments θν and θν are uth associates in D1 and if treatments φµ and φµ are gth associates in D2 then the treatments xνµ and xν µ are said to be (u : g)th associates in D with (u : g) = (0 : 0), (0 : 1), . . . , (0 : m2 ), (1 : 0), . . . , (m1 : m2 ). Proof We shall prove the theorem by establishing that the concordance matrix and the association matrices satisfy certain conditions as established by Bose and Mesner (1959) and as used in Chapter 4. Let B u (u = 0, 1, . . . , m1 ) and B ∗g (g = 0, 1, . . . , m2 ) be the association matrices for designs D1 and D2 , respectively. We then have N 1 N 1

=

m1

λu B u

(5.18)

λ∗g B ∗g

(5.19)

u=0

and N 2 N 2

=

m2 g=0

Then, using (5.16), (5.18), and (5.19), we obtain N N = (N 1 × N 2 )(N 1 × N 2 ) = (N 1 N 1 ) × (N 2 N 2 )

174


 =

m1





λu B u  × 

u=0 m2 m1

=

m2

 λ∗g B ∗g 

g=0

λu λ∗g B u × B ∗g

(5.20)

u=0 g=0

Now let B u:g = B u × B ∗g

(5.21)

It can be seen easily that the B u:g are association matrices since (i) B 0:0 = B × B ∗0 = I t m2 m1 (ii) B u:g = B u × B ∗g u

u=0 g=0

=

g

Bu ×

u

(iii) B u:g B v:h

B ∗g

= It1 It1 × It2 It2 = It It

g

= B u × B ∗g B v × B ∗h = (B u B v ) × B ∗g B ∗h  m m 1 2 f k puv Bk ×  qgh B ∗f  = =

k=0 m2 m 1

f =0

k f puv qgh B k × B ∗f

k=0 f =0

=

m2 m1

f

k puv qgh B k:f

k=0 f =0

=

m2 m1

k:f

pu:g,v:h B k:f

k=0 f =0

with k:f

f

k pu:g,v:h = puv qgh

and k:f

k:f

pu:g,v:h = pu:h,u:g Further, the association scheme defined by the association matrices (5.21) is indeed the same as that stated in the theorem, which is immediately obvious from the way the treatments of N have been enumerated in (5.17). This completes the proof of the theorem.

175


By way of proving the theorem, we also obtain the remaining parameters of the Kronecker product design D as nu:f = nu n∗f λu:f = λu λ∗f k:f P k:f = pu:g,v:h f k = puv · pgh = P k × P ∗f with (u : f ) = (0 : 0), (0 : 1), . . . , (0 : m2 ), (1, 0), . . . , (m1 : m2 ). As a simple application we consider the following example. Example 5.7 Let D1 be a BIB design with parameters t1 = 4, b1 = 4, k1 = 3, r1 = 3(= λ0 ), λ1 = 2, and incidence matrix   1 1 1 0 1 1 0 1  N1 =  1 0 1 1 0 1 1 1 and let D2 be a BIB design with t2 = 3, b2 and incidence matrix  1 1 N 2 = 1 0 0 1

= 3, k2 = 2, r2 = 2(= λ∗0 ), λ∗1 = 1,  0 1 1

Then D has the parameters t = 12, b = 12, k = 6, r = 6(= λ0:0 ), λ0:1 = 3; λ1:0 = 4, λ1:1 = 2, and incidence matrix   N2 N2 N2 0 N 2 N 2 0 N 2   N = N 2 0 N 2 N 2  0 N2 N2 N2 With B 0 = I t1

B 1 = It1 It1 − I t1

B ∗0 = I t2

B ∗1 = It2 It2 − I t2

as the association matrices for D1 and D2 , respectively, the association matrices for D are B 0:0 = I t1 × I t2 = I t B 0:1 = I t1 × It2 It2 − I t1 = I t1 × It2 It2 − I t

176


B 1:0 = It1 It1 − I t1 × I t2 = It1 It1 × I t2 − I t B 1:1 = It1 It1 − I t1 × It2 It2 − I t2 = It It − It1 It1 × I t2 − I t1 × It2 It2 + I t Obviously B 0:0 + B 0:1 + B 1:0 + B 1:1 = It It Any treatment in D is denoted, according to (5.17), by (i, j ) where i = 1, 2, 3, 4, j = 1, 2, 3, and these treatments are arranged in the order as given by (5.17). The association scheme for this PBIB(3) design can then be expressed in two ways, both of which are useful: 1. Any two treatments (i, j ) and (i , j ) are 0:0 0:1 1:0 1:1

associates associates associates associates

if if if if

i i i i

= i = i = i = i

j j j j

= j = j = j = j

or 2. If the treatments are arranged in the 4 × 3 array (1, (2, (3, (4,

1) 1) 1) 1)

(1, (2, (3, (4,

2), 2), 2), 2),

(1, (2, (3, (4,

3) 3) 3) 3)

then each treatment is the 0:0th associate of itself, and any two treatments are 0:1 associates if they are in the same row 1:0 associates if they are in the same column 1:1 associates if they are in different rows and columns This association scheme has been referred to as the extended group-divisible association scheme for an EGD(3)-PBIB design by Hinkelmann and Kempthorne (1963) using representation 1 above, and as the rectangular association scheme by Vartak (1955) using representation 2 (see Sections 4.6.9 and 4.6.5). Finally, the P matrices for the design D are obtained from those for design D1 : 1 0 0 1 P1 = P0 = 0 3 1 2

177


and design D2 : P ∗0 =

1 0 0 2

P ∗1 =

0 1 1 1

as  P 0:0



1

   =  

φ 2 3 φ

 P 1:0

1

 φ  = 1 0

0 2

0 2 0

      6  0   2 0 4

 P 0:1

0

1

  1 =   φ

1

 φ 0 3

 P 1:1

0

 φ  = 0 1

1 1

1 0 2

    3  3

 1   1 2 2

The method of constructing PBIB designs as Kronecker product designs can, obviously, be extended to Kronecker products of incidence matrices for three or more existing PBIB designs, for example, N = N1 × N2 × N3 and the reader should have no difficulties working out the parameters of the resulting PBIB design with (m1 + 1)(m2 + 1)(m3 + 1) − 1 associate classes, where m1 , m2 , m3 are the numbers of associate classes of the component designs. 5.4.3 Usefulness of Kronecker Product Designs It is, of course, clear that the usefulness of Kronecker product designs is somewhat limited as even for two component designs, that is, for N = N 1 × N 2 they may require rather large block sizes and/or large numbers of replications depending on the corresponding parameters for the underlying designs N 1 and N 2 . In general, therefore, only those designs N 1 and N 2 with moderate values for k and/or r will be useful in this connection. Also, the number of associate classes for Kronecker product PBIB designs may be quite large since the maximum number of associate classes for N = N 1 × N 2 is m1 m2 + m1 + m2 . Vartak (1955) and Kageyama (1972) have shown, however, that under certain conditions this number can be reduced considerably, thereby actually inducing a different association scheme. Kageyama (1972), for example, proves that a necessary and sufficient condition for an EGD-(2ν − 1)-PBIB design (see Section 4.6.9) based on the ν-fold Kronecker product of BIB designs to reduce to PBIB(ν) designs having the hypercubic association scheme (see Section 4.6.10) is that the ν BIB designs have the same number of treatments and the same

178


block size. This not only makes Kronecker product PBIB designs more attractive from a practical point of view but it also provides methods of constructing PBIB designs with different association schemes. 5.5 EXTENDED GROUP-DIVISIBLE PBIB DESIGNS We have mentioned earlier that the EGD-PBIB designs play a major role in connection with systems of confounding for asymmetrical factorial experiments and balanced factorial experiments (see Chapter 12). Knowing how to construct EGD-PBIB designs therefore means knowing how to construct such systems of confounding. It is for this reason that we shall present a few methods of constructing EGD-PBIB designs. 5.5.1 EGD-PBIB Designs as Kronecker Product Designs As illustrated already by Example 5.7, the EGD-PBIB(3) design can be constructed as the Kronecker product of two BIB designs. It is then quite obvious that a ν-fold Kronecker product of BIB designs will lead to an EGD-PBIB(2ν − 1) design. More formally, we state the following theorem. Theorem 5.5 Let there be ν designs Di with parameters ti , bi , ki , ri , λ(i) and incidence matrix N i (i = 1, 2, . . . , ν), respectively. Then the Kronecker product design D as given by its incidence matrix N = N1 × N2 × · · · × Nν is an EGD/(2ν − 1)-PBIB design with parameters t=

ν

ti

b=

i=1

λγ1 γ2 ,...,γν =

ν

bi

k=

i=1 ν

ν i=1

ki

r=

ν

ri

i=1

1−γi γi λ(i)

ri

i=1

where the power γi = 0 or 1. Proof For the proof we simply refer to the definition of the Kronecker product design (Section 5.4) and the definition of the EGD/(2ν − 1)-PBIB design (Section 4.6.9). 5.5.2 Method of Balanced Arrays The second method of constructing EGD-PBIB designs is due to Aggarwal (1974). It is based on balanced arrays. Before we give the corresponding theorem we need first the following definition.

EXTENDED GROUP-DIVISIBLE PBIB DESIGNS

179

Definition 5.2 Let α = (α1 , α2 , . . . , αq ) and β = (β1 , β2 , . . . , βq ) be two arrays of q elements, then the symbolic inner product (SIP), α β , of these arrays is the array defined as α β = (α1 β1 , α2 β2 , . . . , αq βq )

We then state the following theorem. Theorem existence of an EGD/(2ν −1)-PBIB design with paramν 5.6 The eters t = i=1 ti , b = νi=2 ti (ti − 1), k = ti , r = νi=2 (ti − 1), and λγ1 γ2 ...γν =

1 if γ1 = γ2 = . . . = γν = 1 0 otherwise

where t1 ≤ ti (i = 2, 3, . . . , ν), is implied by the existence of ν − 1 balanced arrays BAi [ti (ti − 1), t1 , ti , 2](i = 2, 3, . . . , ν), that is, with Ni = ti (ti − 1) assemblies, t1 constraints, strength 2, ti levels, and for each i 0 if x1 = x2 λ(x1 , x2 ) = 1 otherwise [for a definition of balanced array (BA) see Appendix C]. Proof This proof actually gives the construction of the EGD-PBIB design, assuming that the BAi exist. Denote the columns of BAi by a i1 , a i2 , . . . , a i,(ti −1) and the elements in BAi by 0, 1, . . . , ti − 1. Then construct successively the following EGD-PBIB designs: 1. Take the SIP of the array [0, 1, 2, . . . , (t1 − 1)] and each a 2j [j = 1, 2, . . . , t2 (t2 − 1)] of BA2 . If each SIP is taken as a block with treatments of the form (x1 x2 ), then this yields an EGD(3)-PBIB design with the following parameters: t = t1 t2 , b = t2 (t2 − 1), k = t1 , r = t2 − 1, n01 = t2 − 1, n10 = t1 − 1, n11 = (t1 − 1)(t2 − 1), λ01 = 0, λ10 = 0, λ11 = 1. This can be verified easily by noting that (a) each element of BA2 must occur t2 − 1 times in each row of BA2 , hence each element in (0, 1, . . . , t1 − 1) is combined with each element of (0, 1, . . . , t2 − 1), implying r = t2 − 1; (b) each SIP forms a block of size k = t1 and hence b = t2 (t2 − 1); (c) because of the BA property concerning the λ(x1 , x2 ), no two pairs of elements, that is, no two treatments (i, j ) and (i , j ), say, occur together in the same block if i = i or j = j , implying λ01 = λ10 = 0, and two pairs of elements (i, j ) and (i , j ) occur together exactly once if i = i and j = j , implying λ11 = 1. This establishes the EGD(3)-PBIB property of this design. We denote the blocks, that is, SIP arrays, of this EGD-PBIB design by b21 , b22 , . . . , b2,t2 (t2 −1) .

180


2. Then take the SIP of the b2j [j = 1, 2, . . . , t2 (t2 − 1)] with the a 3 [ = 1, 2, . . . , t3 (t3 − 1)] of BA3 . This leads to an EGD/7-PBIB design with parameters t = t1 t2 t3 , b = t2 (t2 − 1)t3 (t3 − 1), k = t1 , r = (t2 − 1) (t3 − 1), λ001 = λ010 = λ011 = λ100 = λ101 = λ110 = 0, λ111 = 1, n100 = t1 − 1, n010 = t2 − 1, n001 = t3 − 1, n110 = (t1 − 1)(t2 − 1), n101 = (t1 − 1)(t3 − 1), n011 = (t2 − 1)(t3 − 1), n111 = (t1 − 1)(t2 − 1)(t3 − 1). This can be verified by repeating the arguments given in item 1. 3. Continue the process of forming the SIP of the blocks of the EGD/(2q − 1)PBIB design with the columns of BAq+1 until q = ν − 1. The resulting design is then an EGD/(2ν − 1)-PBIB design with parameters as stated in the theorem. Example 5.8 Let t = 36 = 3 × 3 × 4, that is, t1 = 3, t2 = 3, t3 = 4. The two required BAs are (see Appendix C for construction)   0 1 2 0 1 2 BA2 =  2 0 1 1 2 0  1 2 0 2 0 1 

0 1 BA3 =  1 0 2 3

2 3 0 1 3 2 2 3 0 1 3 2

2 3 0 0 1 3 1 0 1

 1 2 3 2 1 0  0 3 2

The SIP of (0, 1, 2) with the columns of BA2 yields b21 = (00, 12, 21) b22 = (01, 10, 22) b23 = (02, 11, 20) b24 = (00, 11, 22) b25 = (01, 12, 20) b26 = (02, 10, 21) The SIP of the b2j ’s with the columns of BA3 yields the final plan with 72 blocks of size 3 as given in Table 5.1.

5.5.3 Direct Method The methods described in Sections ability of appropriate BIB designs described in this section, which is a direct method in the sense that it operations.

5.5.1 and 5.5.2 depend both on the availor BAs, respectively. The method to be due to Chang and Hinkelmann (1987), is is based ab initio on simple combinatorial

181


Table 5.1 EGD-PBIB Design for t = 36, b = 72, k = 3, and r = 6 (000, (001, (002, (003, (000, (001, (002, (003, (000, (001, (002, (003,

121, 120, 123, 122, 122, 123, 120, 121, 123, 122, 121, 120,

212) 213) 210) 211) 213) 212) 211) 210) 211) 210) 213) 212)

(010, (011, (012, (013, (010, (011, (012, (013, (010, (011, (012, (013,

101, 100, 103, 102, 102, 103, 100, 101, 103, 102, 101, 100,

222) 223) 220) 221) 223) 222) 221) 220) 221) 220) 223) 222)

(020, (021, (022, (023, (020, (021, (022, (023, (020, (021, (022, (023,

111, 110, 113, 112, 112, 113, 110, 111, 113, 112, 111, 110,

202) 203) 200) 201) 203) 202) 201) 200) 201) 200) 203) 202)

(000, (001, (002, (003, (000, (001, (002, (003, (000, (001, (002, (003,

111, 110, 113, 112, 112, 113, 110, 111, 113, 112, 111, 110,

222) 223) 220) 221) 223) 222) 221) 220) 221) 220) 223) 222)

(010, (011, (012, (013, (010, (011, (012, (013, (010, (011, (012, (013,

121, 120, 123, 122, 122, 123, 120, 121, 123, 122, 121, 120,

202) 203) 200) 201) 203) 202) 201) 200) 201) 200) 203) 202)

(020, (021, (022, (023, (020, (021, (022, (023, (020, (021, (022, (023,

101, 100, 103, 102, 102, 103, 100, 101, 103, 102, 101, 100,

212) 213) 210) 211) 213) 212) 211) 210) 211) 210) 213) 212)

Before we state the main result, we introduce, for ease of notation, the following definition. Definition 5.3 For x and y intergers and x > y, let x (y) =

x! = x(x − 1) · · · (x − y + 1) (x − y)!

(5.22)

182


We then have the following theorem. Theorem 5.7 For t = νi=1 ti treatments with t1 ≤ t2 ≤ · · · ≤ tν there exists an EGD/(2ν − 1)-PBIB design with parameters t=

ν

b=

ti

i=1

r=

ν

ν

ti(t1 )

i=2

(ti − 1)(t1 −1)

k = t1

i=2

λ(0, 0, . . . , 0) = r

λ(1, . . . , 1) =

ν

(t1 − 2)(t1 −2)

i=2

all other λ(γ ) = 0 The proof is rather elementary but quite lengthy. We shall not go through all the details here and instead refer the reader to Chang and Hinkelmann (1987). Just as for Theorem 5.6, however, the proof is based on the actual construction of the design, and the individual steps of the method will be described now. To do so, it is useful to introduce some notation. Let (x1 , x2 , . . . , xν ) denote a treatment with xi {1, 2, . . . , ti } for i = 1, 2, . . . , ν and xi is called the ith component of the treatment (see Section 4.6.9). We then define a matrix 

P (1)



 (2)  P    P = .   .   . 

(5.23)

P (t1 )

ν−1 × tν with the submatrix P () of dimension t i=2 i × ν−1 ν−1 ∗ and i=2 by ). The tν . (In what follows we shall abbreviate i=1 by () elements of P are the treatments (x1 , x2 , . . . , xν ), and P has the following properties: of dimension

ν−1 i=1 ti

1. The first component of any element is for = 1, 2, . . . , t1 . 2. The last component of any element in the sth column is s (1 ≤ s ≤ tν ). 3. The remaining components of any element are changed from one row to the next by first changing the (ν − 1)th component from 1 through tν−1 , then the (ν − 2)th component from 1 through tν−2 , and so on.

183


For example, for ν = 3 the submatrix P () is of the form   (, 1, 1) (, 1, 2) . . . (, 1, t3 )    (, 2, 1) (, 2, 2) . . . (, 2, t3 )      ... ...   (, t2 , 1) (, t2 , 2) . . . (, t2 , t3 )

(5.24)

for = 1, 2, . . . , t1 . The next step is to define a matrix derived from P ,   q1    q2    Q(P ) =  .   ..    q t1

(5.25)

of order t1 × tν such that q is a row vector from P () ( = 1, 2, . . . , t1 ) and any two vectors, q () and q ( ) , do not have the same row number in P () and P ( ) , respectively. Denote the possible choices of Q(P ) by Q1 , Q2 , . . . , QN (suppressing P ), where N = ∗ ti(t1 ) Example 5.9 Let t = 24, t1 = 2, t2 = 3, t3 = 4. The P matrices as given by (5.24) are   (1, 1, 1) (1, 1, 2) (1, 1, 3) (1, 1, 4) P (1) = (1, 2, 1) (1, 2, 2) (1, 2, 3) (1, 2, 4) (1, 3, 1) (1, 3, 2) (1, 3, 3) (1, 3, 4)   (2, 1, 1) (2, 1, 2) (2, 1, 3) (2, 1, 4) P (2) = (2, 2, 1) (2, 2, 2) (2, 2, 3) (2, 2, 4) (2, 3, 1) (2, 3, 2) (2, 3, 3) (2, 3, 4) With N = 3(2) = 6 we then have (1, 1, 1) Q1 = (2, 2, 1) (1, 1, 1) Q2 = (2, 3, 1) (1, 2, 1) Q3 = (2, 1, 1)

(1, 1, 2) (1, 1, 3) (1, 1, 4) (2, 2, 2) (2, 2, 3) (2, 2, 4) (1, 1, 2) (1, 1, 3) (1, 1, 4) (2, 3, 2) (2, 3, 3) (2, 3, 4) (1, 2, 2) (1, 2, 3) (1, 2, 4) (2, 1, 2) (2, 1, 3) (2, 1, 4)

184


(1, 2, 1) (1, 2, 2) Q4 = (2, 3, 1) (2, 3, 2) (1, 3, 1) (1, 3, 2) Q5 = (2, 1, 1) (2, 1, 2) (1, 3, 1) (1, 3, 2) Q6 = (2, 2, 1) (2, 2, 2)

(1, 2, 3) (1, 2, 4) (2, 3, 3) (2, 3, 4) (1, 3, 3) (1, 3, 4) (2, 1, 3) (2, 1, 4) (1, 3, 3) (1, 3, 4) (2, 2, 3) (2, 2, 4)

The PBIB design is now obtained by forming, for each Q ( = 1, 3, . . . , N ), arrays B j = bj 1 , bj 2 , . . . , bj t1 (5.26) such that 1. bj i is an element of the ith row of Q . 2. For i = i , bj i and bj i are selected from different columns of Q . Thus for any given , there are tν(t1 ) such B j . The total number of arrays then is b=N

× tν(t1 )

=

∗

ti(t1 )

× tν(t1 )

=

ν

ti(t1 )

(5.27)

i=2

Each array (5.26) represents a block of size t1 and (5.27) gives the total number of blocks. Hence the collection of all B j ’s is the desired EGD/(2ν − 1)-PBIB design. From Q1 we obtain the following 12 B 1j : B 11 : (1, 1, 1) (2, 2, 2)

B 14 :

(1, 1, 2) (2, 2, 1)

B 12 : (1, 1, 1) (2, 2, 3)

B 15 :

(1, 1, 2) (2, 2, 1)

B 13 : (1, 1, 1) (2, 2, 4)

B 16 :

(1, 1, 2) (2, 2, 4)

B 17 : (1, 1, 3) (2, 2, 1)

B 10 :

(1, 4, 2) (2, 2, 1)

B 18 : (1, 1, 3) (2, 2, 2)

B 11 :

(1, 1, 4) (2, 2, 2)

B 19 : (1, 1, 3) (2, 2, 4)

B 12 :

(1, 1, 4) (2, 2, 3)

The other blocks can be constructed in a similar way.

It is, of course, obvious from the design above that only treatments that differ in all three components appear together in the same block. In general, this is assured by the definition of the P matrices in (5.24), the Q matrices in (5.25) and the B arrays in (5.26). Hence, only λ(1, 1, . . . , 1) > 0.

185


Finally, we point out that if t1 = 2 (as in Example 5.9) the method just described yields the same design as Aggarwal’s (1974) method described in Section 5.5.2. 5.5.4 Generalization of the Direct Method As stated in Theorem 5.7 the direct method leads to EGD-PBIB designs with block size k = t1 . Following Chang and Hinkelmann (1987) we shall now show that the theorem can be modified and used to construct certain EGD-PBIB designs with blocks of size k = ti where 2 ≤ i ≤ ν. Suppose we have t = t1 × t2 × · · · × tν treatments with t1 ≤ t2 ≤ · · · ≤ tν . Let A = {i1 , i2 , . . . , in } be a subset of 1, 2, . . . , ν with 2 ≤ n ≤ ν. For ease of notation we take A = {1, 2, . . . , n}, but the reader should have no difficulty replacing in actual applications 1, 2, . . . , n by i1 , i2 , . . . , in . The main result can then be stated as in the following theorem. Theorem 5.8 Consider an experiment with t = νi=1 ti treatments with t1 ≤ t2 ≤ · · · ≤ tν . Let A = {1, 2, . . . , n}. Then there exists an EGD/(2ν−1 )-PBIB design with parameters   ν n ν (t ) t = ti b= ti 1  tj  i=1

r=

n

j =n+1

i=2

(ti − 1)(t1 −1)

k = t1

i=2

λ(1, 1, . . . , 1, 0, . . . , 0) = n

n

(ti − 2)(t1 −2)

i=2

all other λ(γ ) = 0 Before we prove the theorem we introduce the notation in the following definition. Definition 5.4 Let a i = (1, 2, . . . , ti ) be 1 × ti vectors for i = 1, 2, . . . , ν. The symbolic direct multiplication (SDM) of a p and a q (1 ≤ p, q ≤ ν), denoted by a p a q is given by an array of tq vectors:       (1, 2) (1, tq ) (1, 1)        (2, 2)   (2, tq )   (2, 1)              ap aq =   ..  ,  ..  , . . . ,  ..   .   .   .        (tp , 1)

(tp , 2)

(tp , tq )

186


where each vector has tp elements and the second components in a given vector are identical and equal to 1, 2, . . . , tq , respectively. n Proof n of Theorem 5.8 We first construct an EGD/(2 − 1)-PBIB design for = i=1 ti treatments according to Theorem 5.7. Let the parameters of this design be b∗ , r ∗ , k ∗ , λ∗ say, as given in Theorem 5.7. Let B ∗j (j = 1, 2, . . . , b∗ ) be a block in the EGD/(2n − 1)-PBIB design. Further, let d i = (1, 2, . . . , ti )(i = 1, 2, . . . , ν), and let A = {1, 2, . . . , ν} − A = {n + 1, n + 2, . . . , ν} (because we have chosen A = {1, 2, . . . , n}). We then define the SDP

t∗

D A = d n+1 ⊗ d n+2 ⊗ · · · ⊗ d ν and consider the SDM B j = B ∗j D A for j = 1, 2, . . . , b∗ . Each B ∗j consists of νj =n+1 tj vectors, each vector having t1 elements, each element being an ν-tuple and representing a treatment. We thus have   ν n ν (t ) b = b∗ × tj = t 1  tj  i

j =n+1

i=2

j =n+1

vectors and each such vector is taken as a block for the EGD/(2ν − 1)-PBIB design. From the construction it is obvious that any two elements in the same block (vector) of B j are different in the first n positions and identical in the last (ν − n) positions. Hence λ(1, . . . , 1, 0, . . . , 0) = λ∗ (1, . . . , 1) and all other λ(γ ) are zero. The other properties of the EGD/(2ν − 1)-PBIB design follow easily. We shall illustrate this procedure to construct a design with blocks of size 3 for t = 24 treatments. Example 5.10 Let t = 24, t1 = 2, t2 = 3, t3 = 4. To construct an EGD-PBIB design with blocks of size k = 3, we take A = {2, 3} and construct an EGD/3∗ PBIB design ν−1 for t = 12 = 3 × 4 according to Theorem 5.7. For this case we define i=2 ti (for ν = 2) to be 1, so that the matrix P of (5.23) becomes   (1, 1) (1, 2) (1, 3) (1, 4)   P = (2, 1) (2, 2) (2, 3) (2, 4) (3, 1) (3, 2) (3, 3) (3, 4) and Q(P ) = P . Then the B ∗j are of the       (1, 1) (1, 1) (1, 1)       (2, 2) (2, 2) (2, 3) (3, 3) (3, 4) (3, 2)

form   (1, 1)   (2, 3) (3, 4)



 (1, 1)   (2, 4) (3, 2)

  (1, 1)   (2, 4) (3, 3)

187

HYPERCUBIC PBIB DESIGNS

plus 18 more blocks. Now A = {1} and hence 1 D {1} = 2 Then, for any B ∗j above we obtain B j = B ∗j D {1} that is, the B j ’s are of the form (1, 1, 1) (1, 1, 2) (1, 1, 1) (1, 1, 2) (2, 2, 1) (2, 2, 2) (2, 2, 1) (2, 2, 2) (3, 3, 1) (3, 3, 2) (3, 4, 1) (3, 4, 2)

···

and we have a total of 48 blocks. All we need to do now to obtain the final result is to interchange the elements in each triplet above so that they are in the “right” order, that is, (1, 1, 1) (2, 1, 1) (1, 1, 1) (2, 1, 2) (1, 2, 2) (2, 2, 2) (1, 2, 2) (2, 2, 2) (1, 3, 3) (2, 3, 3) (1, 3, 4) (2, 3, 4)

···

The parameters for this design then are t = 24, b = 48, k = 3, r = 6, and λ(0, 1, 1) = 1. We conclude this section to point out that the EGD-PBIB designs thus constructed are disconnected except when A = {1, 2, . . . , ν} which, of course, is the design of Theorem 5.7. Though disconnectedness is not desirable in general, it can be useful in connection with asymmetrical factorial experiments (see Chapter 12).

5.6 HYPERCUBIC PBIB DESIGNS Using the result due to Kageyama (1972) that an EGD/(2ν − 1)-PBIB design constructed as a ν-fold Kronecker product of BIB designs reduces to a hypercubic PBIB design with ν associate classes if and only if the ν BIB designs have the same number of treatments and the same block size, we can state the following corollary to Theorem 5.5. Corollary 5.1 Let there be ν BIB designs D i with parameters ti , bi , ki , ri , λ(i) and incidence matrix N i such that ti = q and ki = k ∗ (i = 1, 2, . . . , ν). Then the Kronecker product design D given by its incidence matrix N = N1 × N2 × · · · × Nν

188


is a hypercubic PBIB(ν) design with parameters t = q ν , b = νi=1 bi , k = j (k ∗ )ν , r = νi=1 ri , and λj = i=1 ri νi=j +1 λ(i) for j = 1, 2, . . . , ν. Proof We refer simply to the definition of the hypercubic association scheme (Section 4.6.10). Concerning the expression for λj , we note that because the ti ’s and ki ’s are identical it follows that ri /λ(i) = constant for each i and hence all the λ(γ1 , γ2 , . . . , γν ) of the EDG-PBIB with the same number of unity components are identical. Similarly, for a second method of constructing hypercubic PBIB designs, we consider a special case of Theorem 5.6 and state the following corollary. Corollary 5.2 (Aggarwal, 1974) A hypercubic PBIB design with the parameters t = q ν , b = q ν−1 (q − 1)ν−1 , k = q, r = (q − 1)ν−1 , λ1 = λ2 = · · · = λν−1 = 0, λν = 1 can be constructed if q is a prime or a prime power. For the proof we refer to the proof of Theorem 5.6 and the definition of the hypercubic association scheme (see Section 4.6.10). For another method of constructing hypercubic designs, see Chang (1989).

CHAPTER 6

More Block Designs and Blocking Structures

6.1 INTRODUCTION In Chapters 3 and 5 we described the construction of two rather large classes of incomplete block designs with desirable statistical and mathematical properties. Yet, these designs do not cover the need for incomplete block designs for many practical applications, such as variety, environmental, medical, sensory, and screening trials. It is impossible to create a catalogue with all possible designs because of the large number of combinations of design parameters. Hence there is a need for simple methods of constructing designs. Of particular interest often are resolvable designs (see Section 2.7.2) as they allow a certain amount of flexibility with respect to spatial and/or sequential experimentation. In this chapter we shall present three quite general algorithms to construct resolvable incomplete block designs allowing equal or unequal block sizes. Although in general comparisons among all treatments are of primary interest, there are situations where this is only of secondary importance. In such situations the forms of the experiment is on comparing so-called test treatments versus a control (the word control is used here in a general sense with different meaning in different areas of practical applications). Any of the designs discussed in previous chapters can be used for that purpose by simply declaring one of the t treatments as the control. Because of the special nature of the desired inference, however, special designs have been developed, and we shall discuss them briefly here. So far we have only considered designs with one blocking factor. In most practical situations this is clearly sufficient to develop an efficient experimental protocol. There are, however, situations where it is useful to take into account additional blocking factors to further reduce experimental error. We shall consider Design and Analysis of Experiments. Volume 2: Advanced Experimental Design By Klaus Hinkelmann and Oscar Kempthorne ISBN 0-471-55177-5 Copyright  2005 John Wiley & Sons, Inc.

189

190

MORE BLOCK DESIGNS AND BLOCKING STRUCTURES

here briefly one additional blocking factor, one that leads us to the notion of row–column designs as generalization of the Latin-square-type designs of Chapter I.10. 6.2 ALPHA DESIGNS These designs, called α-designs, were introduced by Patterson and Williams (1976) and further developed by John and Williams (1995) to be used mainly in the setting of variety trials in agronomy. 6.2.1 Construction Method The general ideas and steps of constructing an α-design are as follows: 1. Denote the t treatments by 0, 1, . . . , t − 1. 2. Let t = ks, where k is the desired block size and s is the number of blocks in each replicate, there being r replicates in the final design. 3. Select as the generating array a k × r array α with elements a(p, q) in the set of residues mod s (p = 0, 1, 2, . . . , k − 1; q = 1, 2, . . . , r). 4. Develop each column of α cyclically mod s to generate an intermediate k × rs array, α ∗ . 5. Add i · s to each element in the (i + 1)th row of α ∗ (i = 1, 2, . . . , k − 1) to obtain the final array, α ∗∗ , containing elements 0, 1, 2, . . . , t − 1. 6. Separate α ∗∗ into r sets, where the j th set is formed by the columns, (j − 1)s + 1, (j − 1)s + 2, . . . , (j − 1)s + s(j = 1, 2, . . . , r), each column representing a block. Example 6.1 We shall illustrate this construction method by adapting an example from John and Williams (1995) for t = 12 = 4 × 3; that is, k = 4, s = 3. In Table 6.1 we give the generating array α, for r = 2, the intermediate array α ∗ , the final array α ∗∗ , and the final design. The generating array α in Table 6.1 is called a reduced array, having zeros in the first row and first column. All arrays can be represented in the reduced form by adding suitable elements to individual rows and columns, always reducing mod s. This is convenient for the search of optimal α-designs; that is, designs with the highest efficiency factor E [see (1.108)]. For more details see Williams and John (2000). The efficiency factor itself is closely tied to the concurrence matrix N N . It is easy to verify that N N for the design in Table 6.1 has off-diagonal elements equal to 0, 1, 2, implying that some treatment pairs never occur together in a block, other pairs occur together once, and still others occur together twice. To indicate this property the design will be called a α(0, 1, 2)-design. Just as, in general, 2-associate class PBIB designs are more efficient than 3-associate class PBIB designs for the same parameters (t, b, k, r), so are 2-occurrence class

191

ALPHA DESIGNS

Table 6.1 Construction of Design for t = 12, s = 3, k = 4, and r = 2 α Array 0 0 0 0

0 1 2 0 α ∗ Array

0 0 0 0

1 1 1 1

2 2 2 2

0 1 2 0

1 2 0 1

2 0 1 2

1 5 6 10

2 3 7 11

α ∗∗ Array 0 3 6 9

1 4 7 10

2 5 8 11

0 4 8 9

Final Design Replicate Block

1

2

1

2

3

1

2

3

0 3 6 9

1 4 7 10

2 5 8 11

0 4 8 9

1 5 6 10

2 3 7 11

α-designs preferred over 3-occurrence class α designs (Patterson and Williams, 1976). For example, the generating array 0

0

0

2

0

1

0

1

192


leads to an α(0, 1)-design. The efficiency factor for this design can be computed to be .71, which, in this case, is about the same as that for the α(0, 1, 2)-design. For purposes of comparison we mention that the upper bound (1.110) for the efficiency factor is .82. 6.2.2 Available Software Williams and Talbot (1993) have provided a design generation software package, called ALPHA+ , which produces α-designs in the parameter ranges 2 ≤ r ≤ 10

2 ≤ k ≤ 20

t ≤ 500

A more general software package, called CyDesigN, has been provided by Whitaker, Williams, and John (1997). Both packages are able to construct designs with high efficiency factors relative to the upper bounds for such design thus leading to near-optimal designs. 6.2.3 Alpha Designs with Unequal Block Sizes The construction of α-designs is quite flexible and able to accommodate most any desirable, practical block size for a given t, or to deal with the situation when t cannot be written as t = ks(s > 1). In either case we may write t = s1 k1 + s2 k2 and construct a design with s1 blocks of size k1 and s2 blocks of size k2 . Of particular interest is the case where k2 = k1 − 1, that is, where the block sizes are not too different, differing only by one unit. For this situation an appropriate design can be derived from an α-design as follows: 1. Construct an α-design for t + s2 treatments in s = s1 + s2 blocks of size k1 in each replicate. 2. Delete from this design the treatments labeled t, t + 1, . . . , t + s2 − 1. Example 6.2 Consider t = 11 = 2 × 4 + 1 × 3 that is, s1 = 2, s2 = 1, k1 = 4, k2 = 3. The α-design in Table 6.1 is a design for t + s2 = 11 + 1 = 12 treatments in blocks of size 4. Deleting treatment 11 yields the derived design with two blocks of size 4 and one block of size 3.

GENERALIZED CYCLIC INCOMPLETE BLOCK DESIGNS

193

6.3 GENERALIZED CYCLIC INCOMPLETE BLOCK DESIGNS We have seen in Section 5.3 that cyclic PBIB designs are easy to construct and encompass a large class of useful designs. Jarrett and Hall (1978) extended the idea underlying these designs and introduced the notion of generalized cyclic incomplete block designs. Basically, for t = m × n, the designs are obtained by cyclic development of one or more initial blocks, but rather than adding 1 to the element in the initial blocks(s) we now add successively m to each element and reduce mod t. Example 6.3 Consider t = 12 = 3 × 4; that is, m = 3, n = 4. Let the initial blocks be (0, 2, 3, 7)3

and (1, 4, 5, 9)3

where the subscript indicates the incrementing number. Cyclic development then leads to the following design with b = 8, k = 4: (0, 2, 3, 7)

( 1, 4, 5, 9)

(3, 5, 6, 10)

( 4, 7, 8, 0)

(6, 8, 9, 1)

( 7, 10, 11, 3)

(9, 11, 0, 4)

(10, 1, 2, 6)

Inspection shows that r0 = r2 = r3 = r5 = r6 = r8 = r9 = r11 = 3, r1 = r4 = r7 = r10 = 2. The reason for the two different number of replications can be explained as follows. The treatments can be divided into m = 3 groups of n = 4 elements by considering the residue classes mod m, that is, Si = {i, i + m, . . . , i + m(n − 1)} for i = 0, 1, . . . , m − 1. For our example we have S0 = {0, 3, 6, 9} S1 = {1, 4, 7, 10} S2 = {2, 5, 8, 11} It follows that when an initial block is developed the elements in each residue class are equally replicated as each initial block contributes n blocks to the design. Now, the two initial blocks contain three elements each from S0 and S1 , and two elements from S2 . From this argument it follows immediately that in order to construct an equireplicate design we need three initial blocks such that each residual class

194


is represented the same number of times in these three blocks. One possible set of initial blocks is (0, 2, 3, 7)3

(1, 4, 5, 9)3

(6, 8, 10, 11)3

It is easy to select initial blocks to construct an incomplete block design with desired parameters. There is, however, no guarantee that this would lead to a design with a high efficiency factor. One would need a search algorithm to find the “best” design. In general, though, aiming for nearly equal numbers of concurrences of all pairs of treatments will yield a design with a high efficiency factor. Cyclic development can also be used to construct incomplete block designs with unequal block sizes by specifying initial blocks of different sizes. 6.4 DESIGNS BASED ON THE SUCCESSIVE DIAGONALIZING METHOD Khare and Federer (1981) have presented several methods of constructing resolvable incomplete block designs for various factorizations of t, the number of treatments. In Section 3.3.4 we have discussed their method when t = K 2 and K is a prime or prime power. Here we shall discuss a few other cases. 6.4.1 Designs for t = Kk For K prime or prime power and k < K, where k is the block size, we can use the method described in Section 3.3.4 for t = K 2 . We then consider only replicates 2, 3, . . . , K + 1 and delete in each replicate the treatments numbered Kk + 1, . . . , K 2 . This leads to a binary design with r = K replications and rK blocks of size k. These designs are also called rectangular lattice designs (see Section 18.11). Example 6.4 For t = 12 = 4 × 3 we use the design given in Example 3.8 and delete treatments 13, 14, 15, 16 from replicates 2, 3, 4, and 5. Example 6.5 For t = 8 = 4 × 2 we can again use the design from Example 3.8 and delete treatments 9, 10, . . . , 16 from replicates 2, 3, 4, and 5 to obtain a design with b = 16, k = 2, and r = 4. 6.4.2 Designs with t = n2 We consider here the situation where t is a square but not a square of a prime number or prime power. The n can be written as n = mps , where ps is the smallest prime in n. The first ps + 1 replicates from the n + 1 replicates obtained by the successive diagonalizing method on n2 treatments in blocks of size n yield an incomplete block design with (0, 1)-concurrences for treatment pairs. For additional replicates one can either repeat some of the original ps + 1 replicates or

195

COMPARING TREATMENTS WITH A CONTROL

develop further replicates continuing with the successive diagonalizing method. In the latter case the numbers of concurrences increase, however, in groups of ps . The above method can also be used to generate a design for t = nk(k < n). We first obtain the design for t = n2 and then delete from replicates 2, . . . , ps + 1 the treatments numbered nk + 1, . . . , n2 . The versatility of this method is illustrated by the following example. Example 6.6 Fichtner (2000) considered the case of t = 400 varieties (treatments) in blocks of size k = 10 with r = 2 replicates for an agronomic trial. In the context discussed above, we have t = 40 × 10. Rather than writing out the 40 × 40 array for replicate 2 (of the generating design) one only needs to write out the 40 × 10 array. Thus the first block would contain the varieties (1, 41, 81, 121, 161, 201, 241, 281, 321, 361) Adding 1, 2, . . . , 39 to each element produces the remaining 39 blocks. From this it is easy to see that the first block in replicate 2 contains the varieties (1, 42, 83, 124, 165, 206, 247, 288, 329, 370) The remaining blocks can then be filled in easily using replicate 1. We thus obtain the following replicate: Block

Treatments

1 2 3 .. .

1 2 3

42 43 44

83 84 85

124 125 126

165 166 167 .. .

206 207 208

247 248 249

288 289 290

329 330 331

370 371 372

39 40

39 40

80 41

81 82

122 123

163 164

204 205

245 246

286 287

327 328

368 369

If a third replicate were desired, the method of successive diagonalizing would yield the initial block as (1, 43, 85, 127, 169, 211, 253, 295, 337, 379) The remaining blocks are obtained by columnwise cyclic substitution (as described in Section 3.3.4) using the replicate given above. For other cases of constructing (0, 1)-concurrence resolvable incomplete block designs, we refer the reader to Khare and Federer (1981), where methods are discussed for t = p 2k , t = n3 , t = n2k . 6.5 COMPARING TREATMENTS WITH A CONTROL In some experimental situations one of the treatments, often referred to as the control or standard, may play a major role in that the focus is not so much the

196


comparisons among all treatments but rather the comparisons of the remaining treatments, often referred to as the test treatments, with the control. The function of the controls might simply be to provide an indication whether the treatments are even worth consideration or to show their superiority to established procedures, such as comparing new drugs against an established medication, or new varieties against an old variety. Obviously, any of the designs we have discussed so far would be suitable if we simply declare one of the t treatments as the control. However, because of the limitation on the inference space the designs may not be a good choice and more specialized designs may be called for that are more efficient for this particular purpose. Some of these designs will be described in the following sections. 6.5.1 Supplemented Balance Suppose we have t test treatments and one control treatment. Let us denote the t + 1 treatments by T0 , T1 , T2 , . . . , Tt , where T0 represents the control. In the context of block designs, a natural way to construct suitable designs for comparing T1 , T2 , . . . , Tt with T0 is to consider a block design for T1 , T2 , . . . , Tt and augment or reinforce each block with q(≥ 1) replications of T0 . This idea was suggested by Cox (1958) using a BIB design as the generating block design. Das (1958) referred to such designs as reinforced BIB designs. A more general class of designs, referred to as supplemented balance designs, was introduced by Hoblyn, Pearce, and Freeman (1954) and more formally by Pearce (1960). These designs are characterized by the following form of the concurrence matrix:   s0 λ0 . . . λ0 ..   λ0 s ... . NN =  (6.1) λ1 .  .  ..  λ1 . . .  .. s λ0 implying that the self-concurrence is the same for all the test treatments. Further, T0 occurs together λ0 times with each of T1 , T2 , . . . , Tt , and Ti and Tj (i = j , i, j = 0) occur together λ1 times. The example below of a supplemented balance design was given by Pearce (1953) for t = 4, b = 4, k = 7, r0 = 8, r = 5, λ0 = 10, λ1 = 6: 0 0 0 0

0 0 0 0

1 1 1 1

2 2 2 2

3 3 3 3

4 4 4 4

1 2 3 4

where each row represents a block. We note here in passing that the design above is an example of what we have called an extended block design (see Section I.9.8.5). And for the reinforced BIB design we have λ0 = qr.

197


For incomplete blocks an obvious way to construct a supplemented balance design is to augment each block in a BIB design with t treatments by one or more replications of the control. Under certain conditions such designs are A-optimal (see Hedayat, Jacroux, and Majumdar, 1988; Stufken, 1987). 6.5.2 Efficiencies and Optimality Criteria Clearly, there exist many ways to construct suitable augmented designs for given parameters t, b, and k. It is, therefore, important to distinguish between competing designs and choose the one that is “best” for the purpose at hand, which may not only depend on statistical but also on practical considerations. From the statistical, that is, inferential, point of view we would like to obtain maximum information with respect to comparing T0 versus Ti (i = 1, 2, . . . , t). To formulate this more precisely, we consider the usual model for observations from a block design, that is, yij = µ + βi + τj + eij

(6.2)

y = µI + Xβ β + Xτ τ j + e

(6.3)

or, in matrix notation

The information matrix [see (1.9)], assuming equal block sizes, is given by 1 C = R − NN k

(6.4)

where R = X τ X τ = diag(r0 , r1 , . . . , rt ) and N = Xτ Xβ . We know [see (1.20)] that for an estimable function c τ of the treatment effects var(c τˆ ) = c C − c σe2 For the designs discussed in this section the estimable functions of primary interest are of the form τj − τ0 (j = 1, 2, . . . , t). If we denote by 

−1 −1   P = −1   −1

 1 0 ... ... ... 0 0 1 0 . . . . . . 0  0 0 1 0 . . . 0  = (−1I, I t )  ..  . 0 0 ... ... 0 1

(6.5)

the coefficient matrix for the above contrasts, then the variance–covariance matrix for the estimators for those contrasts is given by P C − P σe2 .

198


If we write C of (6.4) as

c1 C 11

c C= 0 c1

(6.6)

then a generalized inverse of C is given by   0 ··· 0  . C − =  ..  −1 0 C 11

(6.7)

which is equivalent to solving the reduced NE C τˆ = Q [see (1.8)] by imposing the condition τˆ0 = 0 (recall that SAS PROC GLM, e.g., obtains a generalized inverse equivalently, by assuming τˆt = 0 rather than τˆ0 = 0). Using C − of the form (6.7) implies that P C − P = C −1 11 and further that (1) the information matrix for treatment control contrasts is given ii by C 11 of (6.6) [see also Majumdar (1996)] and (2) with C −1 11 = (c ) av. var (τˆj − τˆ0 ) =

t t 1 jj 2 1 1 2 c σe = σ t t dj e j =1

(6.8)

j =1

where dj (j = 1, 2, . . . , t) are the eigenvalues of C 11 [see (1.106) and Constantine (1983)]. For a CRD with rj (j = 0, 1, 2, . . . , t) replications for the j th treatment we have, for j = 0, 1 1 2 + (6.9) σe(CRD) av. var (τˆj − τˆ0 ) = r0 rh where t 1 1 1 = rh t rj j =1

2 can then be defined The ratio of (6.9) over (6.8) assuming σe2 = σe(CRD) as the efficiency factor E for a treatment control comparison design (see also Section 1.12.2), that is,

E=

1/r0 + 1/r h t (1/t) (1/dj ) j =1

can be used to compare competing designs.

(6.10)

199


Another way to compare designs is to make use of (6.8) and find, for a given set of parameters t, b, k, the design D ∗ , say, with minimum value of t 1 1 t dj j =1

where D ∗ is referred to as an A-optimal design (see Hedayat, Jacroux, and Majumdar, 1988, and Section 1.13). An alternative criterion, called MV-optimality, is to minimize max (1/dj )

1≤j ≤t

for competing designs (see Hedayat, Jacroux and, Majumdar, 1988). A different approach to optimal designs is suggested and explored by Bechhofer and Tamhane (1981, 1985). They propose to maximize the confidence coefficient for simultaneous confidence intervals for all τj − τ0 (j = 1, 2, . . . , t), thereby linking the designs to Dunnett’s procedures (see I.7.5.7) for comparing treatments with a control (Dunnett, 1955, 1964). In addition to assuming model (6.2), this method has to make distributional assumptions about the eij ’s, which for practical purposes amounts to assuming normality (see Section 6.5.5). 6.5.3 Balanced Treatment Incomplete Block Designs An important and useful class of treatment control designs was introduced by Bechhofer and Tamhane (1981), a class they refer to as balanced treatment incomplete block designs (BTIBD). 6.5.3.1 Definition and Properties Definition 6.1 A treatment control design with t test treatments and one control treatment in b blocks of size k < t + 1 is called a balanced treatment incomplete block design, denoted by BTIBD(t, b, k; λ0 , λ1 ), if (1) each test treatment occurs together with the control λ0 times in a block and (2) any two test treatments occur together λ1 times in a block. The concurrence matrix of a BTIBD is then  s0 λ0 . . . λ0 s1 λ1   NN =  λ0 λ1 s2  ..  . ... λ0 λ1 . . .

of the form  . . . λ0 . . . λ1   ..  ... .    .. . λ1  λ1 st

(6.11)

200


Comparing (6.11) and (6.1) shows that some designs with supplemented balance are special cases of BTIBDs. Since the BTIBDs are, by definition, connected designs, it follows that rank(C) = t, where C is given in (6.4). It follows then from (6.4), with N = (nj i ), and (6.11) that 1 2 tλ noi − =0 k k b

r0 −

i=1

and 1 2 λ0 (t − 1)λ1 nj i − − =0 rj − k k k b

(j = 1, 2, . . . , t)

(6.12)

i=1

It follows from (6.12) that rj −

1 2 nj i = [λ0 − (t − 1)λ1 ]/k k i

and hence C 11 of (6.6) has the form C 11 = aI + bII

(6.13)

with a=

λ0 + tλ1 k

b=

−λ1 k

(6.14)

Using (6.13) and (6.14) it is then easy to obtain C −1 11 , which is of the same form as C 11 , say C −1 11 = cI + dII

(6.15)

We find c=

k λ0 + tλ1

and d =

kλ1 λ0 (λ0 + tλ1 )

(6.16)

and hence from our discussion above this implies that var(τˆj − τˆ0 ) =

k(λ0 + λ1 ) 2 σ λ0 (λ0 + tλ1 ) e

(j = 1, 2, . . . , t)

(6.17)

and cov(τˆj − τˆ0 , τˆj − τˆ0 ) =

kλ1 σ2 λ0 (λ0 + tλ1 ) e

(j = j )

(6.18)

201


Expressions (6.17) and (6.18) explain, of course, in what sense these designs are balanced. 6.5.3.2 Construction Methods There exist different methods of constructing BTIB designs, and we shall mention a few without going into all the details. One obvious method, mentioned already in Section 6.5.1, is to augment each block of a BIBD(t, b, k, r, λ1 ) with n copies of the control. This is a special case of a more general class of BTIB designs, which we denote (following Hedayat, Jacroux, and Majumdar, 1988) by BTIBD(t, b, k ∗ ; u, s; λ0 , λ1 ) and which is characterized by the following form of the incidence matrix: nj i = 0 or 1

for j = 1, 2, . . . , t

i = 1, 2, . . . , b

and n01 = n02 = · · · = n0s = u + 1 n0,s+1 = n0,s+2 = · · · = n0b = u The special case mentioned above has s = 0 and hence k ∗ = k + u. Such a design is of rectangular or R-type, and the general form is depicted in Figure 6.1, where D1 represents the BIBD(t, b, k, r, λ1 ). For s > 0 the design is said to be of step or S-type. Its general form is shown in Figure 6.1. One method to construct an S-type BTIB(t, b, k ∗ , λ0 , λ1 ) is to choose BIB designs for D2 and D3 , more specifically D2 = BIBD(t, s, k − 1, r(2) ; λ(2) )

(6.19)

D3 = BIBD(t, b − s, k, r(3) ; λ(3) )

(6.20)

and

so that k ∗ = k + u, λ0 = (u + 1)r(2) + ur(3) , and λ1 = λ(2) + λ(3) . For this particular method of construction we make the following comments: 1. Let us denote by D2∗ and D3∗ the two component designs of the S-type BTIB design (see Fig. 6.1). Then D2∗ and D3∗ with D2 and D3 as given in (6.19) and (6.20), respectively, are themselves BTIB designs. They constitute what Bechhofer and Tamhane (1981) have called generator designs (see Definition 6.2). 2. As a special case of S-type designs we can have u = 0; that is, the blocks in D3∗ do not contain the control, which means that D3∗ = D3 . 3. For u = 0 the design D3 can be a RCBD, possibly with b − s = 1.

202


Figure 6.1

General form of (a) R- and (b) S-type BTIB designs.

4. As a trivial case (but only in the context of BTIB designs) the design D2 may have blocks of size 1 with s = mt; that is, each treatment occurs m times. An important case in this context is k ∗ = 2, m = 1 with D2 = {1, 2, . . . , t} and

D3 =

1 1 t −1 2 3 ... t

(6.21)

(6.22)

Hedayat D2 of (6.21) and D3 of (6.22) by t 1 and Majumdar (1984) denote and t 2, respectively, where t p in general denotes all pt distinct blocks of size p for t treatments.

203


Let us now return to the notion of generator designs (GD) and their function in constructing BTIB designs. Bechhofer and Tamhane (1981) give the following definition. Definition 6.2 For t test treatments and blocks of size k ∗ a generator design is a BTIB design no proper subset of whose blocks forms a BTIB design, and no block of which contains only one of the t + 1 treatments. To illustrate this notion we use the following example. Example 6.7 For t  0 GD1 = 0  1

= 3 and k ∗ = 3 consider  0 0 0 0 λ0 = 2  2 3

  0 0 0 GD2 = 1 1 2   2 3 3   1 GD3 = 2 3   0 0 0 1 GD4 = 0 0 0 2   1 2 3 3

λ1 = 0

λ0 = 2

λ1 = 1

λ0 = 0

λ1 = 1

λ0 = 2

λ1 = 1

Of these four designs GD 1 , GD2 , and GD3 are generator designs, but GD4 is not, because GD4 = GD1 GD3 . The design GD4 illustrates the construction of a BTIB design, D say, in terms of the generator designs: f2 GD2 f3 GD3 D = f1 GD1 (6.23) with fi ≥ 0(i = 1, 2, 3) and at least f1 or f2 ≥ 1 for an implementable BTIB design. We note that the design D in (6.23) with fi ≥ 1(i = 1, 2, 3) represents another form of S-type BTIB design, which we might refer to as a S2 -type design because it has two steps. In general, we might have a Sq -type design D=

q

fi GDi

i=1

with GDi having bi blocks and b = i fi bi . Apart from the control the GDi represent BIB/RCB designs with increasing block size k(1 ≤ k ≤ k ∗ ).

204


The above then represents a method of constructing generator designs. Another method is to start with a BIBD(t ∗ , b, k, r; λ) with t ∗ > t and identify the treatments t + 1, t + 2, . . . , t ∗ with the control, that is, treatment 0 (Bechhofer and Tamhane, 1981). Any block with 0’s only would be deleted. We illustrate this method with the following example. Example 6.8

Starting with the BIBD(7, 7, 4, 4; 2) 3 5 6 7

1 4 6 7

1 2 5 7

1 2 3 6

2 3 4 7

1 3 4 5

2 4 5 6

and replacing 7 by 0 leads to the BTIBD(6, 7, 4; 2, 2): 3 5 6 0

1 4 6 0

1 2 5 0

1 2 3 6

2 3 4 0

1 3 4 5

2 4 5 6

(6.24)

Replacing 6 and 7 by 0 leads to the BTIBD(5, 7, 4; 4, 2): 3 5 0 0

1 4 0 0

1 2 5 0

1 2 3 0

2 3 4 0

1 3 4 5

2 4 5 0

(6.25)

Replacing 5, 6, and 7 by 0 leads to the BTIBD(4, 7, 4; 6, 2): 3 0 0 0

1 4 0 0

1 2 0 0

1 2 3 0

2 3 4 0

1 3 4 0

2 4 0 0

(6.26)

Suppose we want to construct a BTIBD(4, 8, 4; λ0 , λ1 ). We could adjoin to the GD1 given by (6.26) the GD2 = 4 4 and obtain the following BTIBD(4, 8, 4; 6, 3): 3 0 0 0

1 4 0 0

1 2 0 0

1 2 3 0

2 3 4 0

1 3 4 0

2 4 0 0

1 2 3 4

(6.27)

205


We should mention that instead of replacing 5, 6, 7, by 0 we could have replaced, for example, 6, 7, and 2 by 0, yielding 0 0 3 5

0 0 1 4

0 0 1 5

0 0 1 3

0 0 3 4

0 0 4 5

1 3 4 5

Other methods of constructing generator designs are given by Bechhofer and Tamhane (1981). 6.5.4 Partially Balanced Treatment Incomplete Block Designs To further enrich the class of treatment control designs, an obvious step is to generalize BTIB designs to PBTIB, or partially balanced treatment incomplete block designs, just as BIB designs were generalized to PBIB designs. Rashed (1984) introduced two types of PBTIB designs, and we shall give some of his results below. Another form of PBTIB design was defined by Jacroux (1987) (see also Stufken, 1991). 6.5.4.1 Definitions and Structures Definition 6.3 (Rashed, 1984) A treatment control design for t test treatments and one control is said to be a PBTIB type I design if it satisfies the following conditions: 1. The experimental units are divided into b blocks of size k(k ≤ t). 2. Each test treatment occurs in r blocks, but at most once in a block, and the control is replicated r0 times with possibly multiple applications in a block. 3. The set of test treatments T = {1, 2, . . . , t} is divided into two disjoint subsets, T1 and T2 , with q and t − q treatments, respectively. Each test treatment occurs together in a block with the control λ01 or λ02 times, where λ01 =

b

nj i n0i

for

j ∈ T1

nj i n0i

for

j ∈ T2

i=1

λ02 =

b i=1

4. Any two treatments in T occur together in a block λ times ( = 1, 2). 5. Any two treatments, one from T1 and the other from T2 , occur together in a block λ12 times.

206


It follows then that the matrix C 11 of (6.6) is of the form   eIq Iq aI q + bIq Iq  C 11 =  eIq Iq cI q + dIq Iq

(6.28)

with q = t − q and a = r − r/k + λ11 /k, b = −λ11 /k, c = r − r/k + λ22 /k, d = −λ11 /k, and e = −λ12 /k. Since C −1 11 is of the same form as C11 of (6.28) we can state the following theorem. Theorem 6.1 For a PBTIB type I design the variance–covariance structure of the estimators τj − τ0 for τj − τ0 is given by  2 2  α1 σe for j ∈ T1 var( τj − τ0 ) =  α 2 σ 2 for j ∈ T 2 2 e  2   ρ 1 σe cov( τj − τ0 , τj − τ0 ) = ρ2 σe2  ρ σ 2 3 e

for for for

j, i ∈ T1 j, i ∈ T2 j ∈ T1 , j ∈ T2

where α12 , α22 , ρ1 , ρ2 , ρ3 are functions of the design parameters (see Rashed, 1984). An illustration of a PBTIB type I design is given in the following example. Example 6.9 For t = 5, b = 8, k = 3, r = 3, r0 = 9 the two groups are T1 = {2, 3, 4}, T2 = {1, 5} with λ01 = 2, λ02 = 3, λ11 = 1, λ22 = 0, λ12 = 1. The design is given by   0 0 0 0 0 0 1 2   D = 0 0 0 1 3 4 2 3 1 2 5 3 4 5 4 5 and



C −1 11

.619 .19   = .19  .166 .166

.19 .619 .19 .166 .166

.19 .19 .619 .166 .166

.166 .166 .166 .583 .083

 .166 .166   .166  .083 .583

that is, α12 = .619, α22 = .583, ρ1 = .19, ρ2 = .083, and ρ3 = .166.

207


Another type of PBTIB design that retains the feature of two variances but allows for up to six different covariances for comparison estimators is described in the following definition. Definition 6.4 (Rashed, 1984) A treatment control design for t test treatments and one control is said to be a PBTIB type II design if it satisfies the following conditions. Conditions 1, 2, and 3 are as in Definition 6.3: 4. Each group Ti considered by itself has either a BIB or a PBIB association structure, with at least one of these groups having PBIB design properties; (2) that is, any two treatments in Ti occur together in either λ(1) ii or λii blocks with λii (1) = λ(2) ii for at least one i = 1, 2. 5. Any two treatments, one from T1 and the other from T2 , occur together in (2) λ12 blocks with λ12 = λ(1) ii or λii (i = 1, 2). We note that, if in condition 4 above both groups have BIB association structure, then the PBTIB type II design reduces to a PBTIB type I design. This gives an indication how much more general type II designs are compared to type I designs. For type II designs Rashed (1984) proved the following theorem. Theorem 6.2 For a PBTIB type II design the variance–covariance structure of the estimators τj − τ0 for τj − τ0 is given by

τ0 ) = var( τj −

 2 2  α1 σe

for

j ∈ T1

 α 2 σ 2

for

j ∈ T2

2 e

 2  ρ11 σe cov( τj − τ0 , τj − τ0 ) = ρ21 σe2  ρ σ 2 31 e

or or or

ρ12 σe2 ρ22 σe2 ρ32 σe2

for for for

j, j ∈ T1 j, j ∈ T2 j ∈ T1 , j ∈ T2

We shall illustrate this structure with the following example. Example 6.10 For t = 6, b = 8, k = 3, r = 3, r = 3, r0 = 6, the two groups are T1 = {1, 2, 5, 6} and T2 = {3, 4}, and the design is given by   1 2 3 0 0 0 0 0 D = 2 3 4 4 5 1 0 1 4 5 6 5 6 6 2 3

208


(2) with λ01 = 2, λ02 = 1, λ(1) 11 = 1, λ11 = 0 (for a PBIB association structure):

0th Associate

1st Associates

1 2 5 6

2, 1, 2, 1,

2nd Associates

6 5 6 5

5 6 1 2

and λ(1) 22 = 1 (for a BIB association structure) and λ12 = 1. The variance–covariance matrix (if the treatments are written in the order 1, 2, 5, 6, 3, 4) is found to be 

C −1 11

.66 .22  .16 = .22 .25 .25

.22 .66 .22 .16 .25 .25

.16 .22 .66 .22 .25 .25

.22 .16 .22 .66 .25 .25

.25 .25 .25 .25 .71 .29

 .25 .25  .25  .25 .29 .71

that is, α12 = .66, α22 = .71, ρ11 = .22, ρ12 = .16, ρ21 = .20, ρ31 = .25.

6.5.4.2 Construction of PBTIB Designs We shall describe briefly one method of constructing PBTIB designs as developed by Rashed (1984). It is an extension of one of the methods described by Bechhofer and Tamhane (1981) to construct BTIB designs (see Section 6.5.3.2). Generally speaking, in order to construct a PBTIB design with t treatments in b blocks of size k(≤t), we start with a PBIB design with t ∗ (>t) treatments in b∗ (≤b) blocks of size k ∗ = k, and then replace the t ∗ − t “excess” treatments by 0, deleting any resulting blocks which contain only the control. This is a very general recipe and may not necessarily lead to a PBTIB design unless we impose further conditions. Therefore, to be more precise, we start with a GD-PBIB design (see Section 4.6.1) and proceed as follows: 1. Write out the GD association scheme for t ∗ = t1∗ × t2∗ treatments, that is, an array of t1∗ rows and t2∗ columns, which we denote by T . 2. For t = (t11 × t12 ) + (t21 × t22 ) form two nonoverlapping subarrays T1 (t11 × t12 ) and T2 (t21 × t22 ) within T such that T1 and T2 each form a GD association scheme. 3. Denote the remaining set of treatments by T0 and replace those treatments by 0 (these are the excess treatments mentioned above).

209


Schematically this procedure can be depicted as follows: t12 t11

          

T1 T0 t22

t21

         

T2 !

"# t2∗

t1∗

$

This implies that t11 + t21 ≤ t1∗ t12 ≤ t2∗ t22 ≤ t2∗ Following steps 1–3 and returning to the starting PBIB design yields the required PBTIB design in which each treatment is replicated r times and the control is replicated r0 times where r0 = (t1∗ t2∗ − t11 t12 − t21 t22 )r = (t ∗ − t)r Furthermore, the remaining parameters of the PBTIB design are related to those of the PBIB design as follows: λ(1) ii = λ1

λ(2) ii = λ2

(i = 1, 2)

(λ(1) ii exists only if Ti has more than one treatment in the same row of T , and λ(2) ii exists only if Ti has treatments in more than one row of T ), λ0i = ai λ1 + bi λ2

(i = 1, 2)

where ai equals the number of treatments that are replaced by zeros and are 1st associates of a treatment in Ti , that is, ai = t2∗ − ti2 , and bi equals the number of treatments that are replaced by zeros and are 2nd associates of a treatment in Ti , that is, bi = t ∗ − t − (t2∗ − ti2 ) = t ∗ − t − ai (i = 1, 2), and λ12 = λ2 We shall illustrate the construction of a PBTIB design with the following examples.

210


Example 6.11 Suppose t = 8, k = 3. We start with design SR 23 from Clatworthy (1973) with t ∗ = 9, k = 3, r = 3, b∗ = 9, λ1 = 0, λ2 = 1:   1 4 7 1 2 3 1 2 3 2 5 8 5 6 4 6 4 5 3 6 9 9 7 8 8 9 7 and association structure T : 1 4 7 2 5 8 3 6 9 If we replace 9 by 0, that is, T0 = {9}, then

% & 1 4 7 T1 = T2 = 3 6 2 5 8 and the PBTIB design is given  1 4 2 5 3 6

by  7 1 2 3 1 2 3 8 5 6 4 6 4 5 0 0 7 8 8 0 7

(6.29)

The association schemes for T1 and T2 follow from the association scheme for T as 0th Associate

1st Associates

T1

1 2 4 5 7 8

4, 5, 1, 2, 1, 2,

T2

3 6

6 3

7 8 7 8 4 5

2nd Associates 2, 1, 2, 1, 2, 1,

3, 3, 3, 3, 3, 3,

5, 4, 5, 4, 5, 4,

6, 6, 6, 6, 6, 6,

8 7 8 7 8 7

1, 2, 4, 5, 7, 8 1, 2, 4, 5, 7, 8

It follows, therefore, that the design (6.29) is a PBTIB type II design with param(2) (1) eters r = r0 = 3, λ01 = 1, λ02 = 0, λ(1) 11 = 0, λ11 = 1, λ22 = 0, λ12 = 1. Example 6.12 Suppose t = 7, k = 3. We start again with design SR 23 as given in Example 6.11. We now replace 8 and 9 by 0, that is, T0 = {8, 9} and then have

% & 2 5 T1 = 1 4 7 T2 = 3 6

211


or, schematically, 1

4

7

2

5

8

3

6

9

←−

T1

←−

T0

↑ T2 The resulting PBTIB design is   1 4 7 1 2 3 1 2 3   2 5 0 5 6 4 6 4 5

(6.30)

3 6 0 0 7 0 0 0 7 (1) (2) with parameters r0 = 6, λ01 = 2, λ02 = 1, λ(1) 11 = 0, λ22 = 0, λ22 = 1, λ12 = 1. It is easy to verify that the design (6.30) is a PBTIB type II design. Since there exist many GD-PBIB designs (see Clatworthy, 1973) the method described above will yield a rather rich class of PBTIB designs, most of them being of type II.

6.5.5 Optimal Designs In the preceding sections we have described the construction of various types of treatment control designs. For a prespecified value of the number of test treatments t and block size k, different methods may lead to different designs, some more efficient than others. In deciding which design to choose among competing designs and to use in a practical application, one consideration may be to compare their efficiencies and look, if possible, for the optimal design. In this context different optimality criteria have been used (see Section 6.5.2), based on different philosophies [for a discussion see Hedayat, Jacroux, and Majumdar (1988), and the comments following their article]. Unfortunately, different optimality criteria may lead to different optimal designs, but often the optimal design using one criterion is near optimal under another criterion. Using the A-optimality criterion Hedayat and Majumdar (1984) provide for the class of BTIB designs a catalogue of optimal R- and S-type designs for block sizes 2 ≤ k ∗ ≤ 8. Optimal R-type designs are also given by Stufken (1987). These results are based on a theorem by Majumdar and Notz (1983), which can be stated as in the following theorem (Hedayat and Majumdar, 1984). Theorem 6.3 For given (t, b, k ∗ ) a BTIBD(t, b, k ∗ , u, s; λ0 , λ1 ) (see Section 6.5.3.2) is A-optimal if for x = u and z = s the following function is minimized: g(x, z) = (t − 1)2 [btk ∗ (k ∗ − 1) − (bx + z)(k ∗ t − t + k) + (bx 2 + 2xz + z)]−1 + [k ∗ (bx + x) − (bx 2 + 2xz + z)]−1

212


among integers x = 0, 1, . . . , [k ∗ /2], z = 0, 1, . . . , b − 1, with [k ∗ /2] the largest integer not greater than k ∗ /2, and z positive when x = 0 and z = 0 when x = [k ∗ /2]. A different approach to constructing optimal BTIB designs has been provided by Bechhofer and Tamhane (1983a, 1983b, 1985). Their criterion is maximization of the confidence coefficient for the joint confidence intervals of prescribed allowance (or length) for τj − τ0 (j = 1, 2, . . . , t), namely, for > 0, P1 = P {τj − τ0 ≥ τj − τ0 − ; j = 1, 2, . . . , t}

(6.31)

for one-sided confidence intervals, and P2 = P { τj − τ0 − ≤ τj − τ0 ≤ τj − τ0 + ; j = 1, 2, . . . , t}

(6.32)

for two-sided confidence intervals. This requires that additional assumptions have to be made about the error eij in model (6.2). Assuming that the eij are i.i.d. N (0, σe2 ) and denoting the right-hand sides of (6.17) and (6.18) by α 2 σe2 and ρα 2 σe2 , respectively, then expressions for P1 in (6.31) and P2 in (6.32) can be written as ) '∞ ( √ x ρ + /ασe t P1 = d(x) √ 1−ρ

(6.33)

√ ) '∞ ( √ x ρ + /ασe x ρ − /ασe t d(x) P2 = − √ √ 1−ρ 1−ρ

(6.34)

−∞

and

−∞

where (·) denotes the standard normal cumulative distribution function (cdf). We note here, parenthetically, that the assumption about the eij does not agree with our underlying framework of randomization theory (see Sections 1.6 and I.10). We, therefore, consider the confidence coefficients in (6.33) and (6.34) to be approximate confidence coefficients. Nevertheless, the proposed procedures of finding optimal designs based on maximizing P1 or P2 are useful. We note also that for design purposes σe2 is assumed to be known, based on prior experiences or theoretical subject matter considerations. But as Bechhofer and Tamhane (1985) suggest, for analysis purposes an estimate of σe2 , based on the analysis of variance, should be used in connection with Dunnett’s procedure (Dunnett, 1955; see also I.7). Tables of optimal BTIB designs for one-sided and two-sided comparisons for 2 ≤ t ≤ 6), k ∗ = 2, 3 and a wide range of values for b and /σe are given by Bechhofer and Tamhane (1985). The procedure is, for a desired confidence coefficient, to combine one or more replicates from a table of minimal complete sets of generator designs.

213

ROW–COLUMN DESIGNS

6.6 ROW–COLUMN DESIGNS 6.6.1 Introduction For purposes of error control or error reduction the principle of blocking is of primary importance. So far we have considered only one blocking factor, and for most situations this is quite satisfactory. There are, however, situations where blocking in more than one dimension may be needed. For example, in an agronomic or forestry trial it is not uncommon to compare a large number of hybrid varieties (which represent the treatments), which leads to a large experimental area, and it may be important to account for or eliminate the effects of fertility trends in the land in two directions. This implies that one needs to block or eliminate heterogeneity in two directions. The two blocking systems are referred to generally (borrowing terminology from field experiments) as row blocking and column blocking, and the resulting designs are referred to as row–column designs. Examples of such designs are the Latin square type designs of Chapter I.10, but other designs may be needed with incomplete blocks in both directions. As an example consider the following design given by John and Williams (1995) with t = 12 treatments in r = 4 rows and c = 9 columns:

Row

1 2 3 4

1

2

3

4

4 12 7 11

8 11 10 9

5 2 9 12

9 3 4 6

Column 5 6 11 5 6 1

3 10 5 7

7

8

9

2 4 1 10

1 8 12 3

7 6 8 2

Both rows and columns represent incomplete blocks, and the row design and column design, which we refer to as the component designs, are characterized by their respective incidence matrices, N ρ and N γ say. We shall see later (see Section 6.6.2) that the properties of the row–column design depend on N ρ and N γ , but that it is not straightforward how to combine the component designs to produce a “good” final design. We shall return to this point in Section 6.6.3 after discussing the analysis of a row–column design in Section 6.6.2. 6.6.2 Model and Normal Equations Consider a row–column design with r rows, c columns, and t treatments where treatment k is replicated rk times (k = 1, 2, . . . , t). Let yij k denote the observation for the experimental unit in the ith row and j th column to which treatment k has been applied. Based on the assumption of unit treatment additivity (see Section I.10.2.3), a model for yij k can be derived [despite criticism by Srivastava (1993, 1996), and Srivastava and Wang (1998)] and written as yij k = µ + ρi + γj + τk + eij k

(6.35)

214


where ρi (i = 1, 2, . . . , r), γj (j = 1, 2, . . . , c), τk (k = 1, 2, . . . , t) represent the row, effects, respectively, which are defined such that and treatment column, ρi = 0, γj = 0, τk = 0 (see Section I.10.2.3). The eij k represent the combined experimental and observational error. Let us rewrite (6.35) in the usual fashion in matrix notation as follows: y = µI + Xρ ρ + Xγ γ + Xτ τ + e

(6.36)

where ρ, γ , τ represent the row, column, and treatment effect vectors, respectively, and Xρ , Xγ , X τ are known matrices linking the row, column, and treatment effects to the observations. If we write, for example, the observation y such that the first c components are the observation in row 1, the next c from row 2, and so forth, then we have   Ic   Ic   Xρ =  (6.37)  ..   . Ic of dimension rc × r, and 

 Ic I c   Xγ =  . . .  Ic

(6.38)

of dimension rc × c. Further, the relationships between Xρ , X γ , and X τ are Xτ Xρ = N ρ

X τ X γ = N γ

(6.39)

Using (6.37), (6.38), and (6.39), it is then easy to write the NE for model (6.36) as 

rc

 cIr  rI  c r

cIr

rIc

cI r

Ir Ic

Ic Ir

rI c

Nρ

Nγ

    Irc y µ    N ρ   Xρ y  ρ      =  γ  X γ y  N γ     δ τ X y r r

(6.40)

τ

where r = (r1 , r2 , . . . , rt ) and r δ = diag(r1 , r2 , . . . , rt ). Similar NE as follows: to Section 1.3.2 we shall derive from (6.40) the reduced Using ρ i = 0 and γj = 0 (as suggested by ρi = 0 and γj = 0), we obtain from the first equation of (6.40) µ=

G r τ − n n

(6.41)

215

ROW–COLUMN DESIGNS

where we are using the following obvious notation for the right-hand side of (6.40):    Irc y G Xρ y  P   =  X γ y    T Xτ y 

(6.42)

and rc = n. From the second and third set of the NE (6.40) we obtain c ρ = P − c µIr − N ρ τ

(6.43)

τ r γ = − r µIc − N γ

(6.44)

and

Substituting (6.41), (6.43), and (6.44) into the fourth set of equations in (6.40) and simplifying, we obtain the RNE for τ as

1 1 rr 1 G 1 r δ − N ρ N ρ − N γ Nγ + τ = T − NρP − Nγ + r c r n c r n (6.45)

which we write in abbreviated form as C τ =Q

(6.46)

with C and Q as defined in (6.45). A solution to (6.46) can be written as τ = C−Q

(6.47)

where C − is a generalized inverse of C. Any statistical inference about the treatment effects is then derived, similar to the developments in Chapter 1, from (6.46) and (6.47), in conjunction with information from the analysis of variance (see Section 6.6.3). For example, rank (C)(≤t − 1) determines the number of linearly independent estimable functions, τ = c c τ , of the treatment effects. For any estimable function c τ we have c* τ, with τ from (6.47), and var(c τ ) = c C − cσe2 6.6.3 Analysis of Variance Following general principles (see, e.g., Sections I.4.11 and 1.3.6), it is easy to write out the analysis of variance table for a row–column design that will enable

216


Table 6.2 Design

Analysis of Variance for Row–Column

Source

d.f.

Xρ |I

SS r

r −1

c

(y i·· − y ... )2

c−1

c r (y ·j · − y ... )2

i=1

Xγ |I, Xρ

j =1

τ Q

− 1a

Xτ |I, X ρ , X γ

t

I |I, Xρ , Xγ , X τ

Difference ≡ νE

Total

rc − 1

Difference ≡ SS(E) (yij k − y ... )2 i,j,k

a Assuming

a connected design.

us to (i) estimate σe2 and (ii) obtain an (approximate) F test for testing H0 : τ1 = τ2 = · · · = τt = 0. Such an ANOVA table is given in Table 6.2, using an earlier established notation. We then have σe2 = MS(E) =

SS(E) νE

and F =

τ Q/(t − 1) MS(E)

provides an approximate test for testing H0 : τ1 = τ2 = · · · = τt = 0. It should be mentioned (as is, of course, obvious from the notation used) that the ANOVA in Table 6.2 is a sequential (or type I in SAS terminology) ANOVA. The nature of the design implies immediately that SS(X γ |I, Xρ ) = SS(Xγ |I), and hence the simple expression for SS(X γ |I, X ρ ) given in Table 6.2. 6.6.4 An Example We shall illustrate the analysis of data from a row–column design, using the design of Section 6.1 and using SAS PROC GLM. The data and analysis are given in Table 6.3. We make the following comments about the analysis and the SAS output in Table 6.3: 1. The inv-option in the model statement provides a generalized inverse to the coefficient matrix of the NE (6.40), and the 12 × 12 submatrix given by the rows and columns T1 , T2 , . . . , T12 serve as a C − for the RNE (6.46).

217

1 4 9 15 1 8 1 22 2 3 2 20 2 7 4 14 3 2 10 31 3 6 5 35 4 1 11 40 4 5 1 42 4 9 2 39

proc glm data=rowcoll; class R C T; model y = R C T/inv e; lsmeans T/stderr; estimate 'T1-T2' T 1 -1; estimate 'T1-T3' T 1 0 -1; estimate 'T1-T4' T 1 0 0 -1;

proc print data=rowcoll; title1 'TABLE 6.3'; title2 'DATA ROW-COLUMN DESIGN'; title3 'WITH t=12 TREATMENTS, r=4 ROWS, c=9 COLUMNS'; run;

options nodate pageno=1; data rowcoll; input R C T y @@; datalines; 1 1 4 21 1 2 8 18 1 3 5 14 1 5 11 9 1 6 3 17 1 7 2 24 1 9 7 13 2 1 12 8 2 2 11 11 2 4 3 25 2 5 5 17 2 6 10 24 2 8 8 19 2 9 6 25 3 1 7 30 3 3 9 36 3 4 4 29 3 5 6 30 3 7 1 29 3 8 12 28 3 9 8 25 4 2 9 41 4 3 12 43 4 4 6 39 4 6 7 50 4 7 10 47 4 8 3 44 ; run;

Table 6.3 Data and Analysis of Row–Column Design (t = 12, r = 4, c = 9)

218

Table 6.3 (Continued )

R 1 1 1 1 1 1 1 1 1 2 2 2 2 2

Obs 1 2 3 4 5 6 7 8 9 10 11 12 13 14

1 2 3 4 5 6 7 8 9 1 2 3 4 5

C 4 8 5 9 11 3 2 1 7 12 11 2 3 5

T 21 18 14 15 9 17 24 22 13 8 11 20 25 17

Y

estimate 'T1-T5' T 1 0 0 0 -1; estimate 'T1-T6' T 1 0 0 0 0 -1; estimate 'T1-T7' T 1 0 0 0 0 0 -1; estimate 'T1-T8' T 1 0 0 0 0 0 0 -1; estimate 'T1-T9' T 1 0 0 0 0 0 0 0 -1; estimate 'T1-T10' T 1 0 0 0 0 0 0 0 0 -1; estimate 'T1-T11' T 1 0 0 0 0 0 0 0 0 0 -1; estimate 'T1-T12' T 1 0 0 0 0 0 0 0 0 0 0 -1; title2 'ANALYSIS OF VARIANCE'; title3 'FOR ROW-COLUMN DESIGN'; run;

219

15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36

2 2 2 2 3 3 3 3 3 3 3 3 3 4 4 4 4 4 4 4 4 4

6 7 8 9 1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9

10 4 8 6 7 10 9 4 6 5 1 12 8 11 9 12 6 1 7 10 3 2

24 14 19 25 30 31 36 29 30 35 29 28 25 40 41 43 39 42 50 47 44 39

220

Intercept R 1 R 2 R 3 R 4 C 1 C 2 C 3 C 4 C 5 C 6 C 7 C 8 C 9

0.8611111111 -0.083333333 -0.125 -0.125 0 -0.444444444 -0.333333333 -0.444444444 -0.333333333 -0.333333333 -0.333333333 -0.333333333 -0.444444444 0

Intercept


R 1 -0.125 0.125 0.25 0.125 0 -9.73435E-17 -6.94448E-17 -9.88873E-17 -5.45625E-17 -7.38351E-17 -6.80636E-17 -6.84973E-17 -1.07621E-16 0

R 2

R 3 -0.125 0.125 0.125 0.25 0 -8.46741E-17 -6.61156E-17 -8.6178E-17 -5.73645E-17 -5.91767E-17 -6.29509E-17 -6.39636E-17 -8.74456E-17 0

X'X Generalized Inverse (g2)

36

1 2 3 4 1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9 10 11 12

Values

Number of observations

4 9 12

Levels

-0.083333333 0.25 0.125 0.125 0 -5.68289E-19 -6.06814E-18 -3.65495E-19 1.010807E-17 -8.63974E-19 5.078598E-18 5.223935E-18 -1.12406E-17 0

R C T

Class

ANALYSIS OF VARIANCE FOR ROW – COLUMN DESIGN The GLM Procedure Class Level Information

0 0 0 0 0 0 0 0 0 0 0 0 0 0

R 4

-0.444444444 -5.68289E-19 -9.73435E-17 -8.46741E-17 0 0.6666666667 0.3333333333 0.3333333333 0.3333333333 0.3333333333 0.3333333333 0.3333333333 0.3333333333 0

C 1

221

1 2 3 4 5 6 7 8 9 10 11 12

Intercept R 1 R 2 R 3 R 4 C 1 C 2 C 3 C 4 C 5 C 6 C 7

T T T T T T T T T T T T Y

-0.333333333 -6.06814E-18 -6.94448E-17 -6.61156E-17 0 0.3333333333 0.6666666667 0.3333333333 0.3333333333 0.3333333333 0.3333333333 0.3333333333

C 2

-0.421296296 -0.532407407 -0.421296296 -0.37962963 -0.37962963 -0.555555556 -0.532407407 -0.490740741 -0.421296296 -0.444444444 -0.421296296 0 34.208333333

0.0416666667 -1.56588E-16 -1.31958E-16 -0.041666667 -0.041666667 -1.63766E-16 0.0416666667 -0.041666667 0.0416666667 -1.38442E-16 -1.28995E-16 0 -24.16666667

-0.444444444 -3.65495E-19 -9.88873E-17 -8.6178E-17 0 0.3333333333 0.3333333333 0.6666666667 0.3333333333 0.3333333333 0.3333333333 0.3333333333

C 3 -0.333333333 1.010807E-17 -5.45625E-17 -5.73645E-17 0 0.3333333333 0.3333333333 0.3333333333 0.6666666667 0.3333333333 0.3333333333 0.3333333333

C 4

-0.333333333 -8.63974E-19 -7.38351E-17 -5.91767E-17 0 0.3333333333 0.3333333333 0.3333333333 0.3333333333 0.6666666667 0.3333333333 0.3333333333

C 5

-8.68327E-17 0.0416666667 0.0416666667 -0.041666667 -0.041666667 -1.12908E-16 -1.05476E-16 -0.041666667 -8.40883E-17 -8.45796E-17 0.0416666667 0 -12.45833333


-0.041666667 -0.041666667 -0.041666667 -0.083333333 -0.083333333 -2.16298E-17 -0.041666667 -0.083333333 -0.041666667 -2.56719E-17 -0.041666667 0 -25.54166667

-0.333333333 5.078598E-18 -6.80636E-17 -6.29509E-17 0 0.3333333333 0.3333333333 0.3333333333 0.3333333333 0.3333333333 0.6666666667 0.3333333333

C 6

0 0 0 0 0 0 0 0 0 0 0 0 0

-0.333333333 5.223935E-18 -6.84973E-17 -6.39636E-17 0 0.3333333333 0.3333333333 0.3333333333 0.3333333333 0.3333333333 0.3333333333 0.6666666667

C 7

0.1111111111 0.2222222222 0.1111111111 -8.59604E-16 0.1111111111 0.2222222222 0.1111111111 0.2222222222 0.1111111111 0.1111111111 -7.20447E-16 0 3

222

8 9 1 2 3 4 5 6 7 8 9 10 11 12

Intercept R 1 R 2 R 3

C C T T T T T T T T T T T T y

C 2

-0.444444444 -1.12406E-17 -1.07621E-16 -8.74456E-17

C 8

0.3333333333 0 -6.6489E-16 0.1111111111 -6.59544E-16 -7.08048E-16 -6.10521E-16 0.1111111111 0.1111111111 -6.29051E-16 -0.111111111 -0.111111111 -0.111111111 0 1.4444444444


0.3333333333 0 -7.25482E-16 0.1111111111 -0.111111111 -0.111111111 -6.70087E-16 -7.08466E-16 0.1111111111 0.1111111111 -0.111111111 -7.945E-16 -6.75091E-16 0 1.5555555556

C 4

0 0 0 0

C 9

-0.421296296 -0.041666667 0.0416666667 -8.68327E-17

T 1

C 5

-0.532407407 -0.041666667 -1.56588E-16 0.0416666667

T 2

0.3333333333 0 -0.111111111 0.1111111111 -6.80361E-16 -7.21926E-16 -0.111111111 -6.35922E-16 0.1111111111 0.1111111111 -6.44453E-16 -7.4153E-16 -0.111111111 0 0.7777777778


0.3333333333 0 0.1111111111 0.1111111111 0.1111111111 0.1111111111 -7.265E-16 0.2222222222 0.2222222222 0.2222222222 -7.48079E-16 0.1111111111 0.1111111111 0 5

C 3

X’X Generalized Inverse (g2)

-0.421296296 -0.041666667 -1.31958E-16 0.0416666667

T 3

0.3333333333 0 -7.19612E-16 0.1111111111 -0.111111111 -7.52338E-16 -0.111111111 0.1111111111 -7.09929E-16 0.1111111111 -6.91857E-16 -0.111111111 -6.70031E-16 0 6.3333333333

C 6

-0.37962963 -0.083333333 -0.041666667 -0.041666667

T 4

0.3333333333 0 -0.111111111 -6.41164E-16 -7.03875E-16 -0.111111111 -6.53118E-16 0.1111111111 0.1111111111 0.1111111111 -6.73674E-16 -0.111111111 -6.46005E-16 0 2.8888888889

C 7

223

4 1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9 10

T 11 T 12 y

R C C C C C C C C C T T T T T T T T T T

0.1111111111 0 5

C 8

0 0.3333333333 0.3333333333 0.3333333333 0.3333333333 0.3333333333 0.3333333333 0.3333333333 0.6666666667 0 -7.15563E-16 0.2222222222 -7.50593E-16 0.1111111111 0.1111111111 0.2222222222 0.2222222222 0.1111111111 0.1111111111 0.1111111111

0 0.1111111111 -6.6489E-16 0.1111111111 -7.25482E-16 -0.1111111111 -7.19612E-16 -0.111111111 -7.15563E-16 0 0.8425925926 0.4212962963 0.4212962963 0.4212962963 0.4212962963 0.4444444444 0.3981481481 0.4212962963 0.3981481481 0.4444444444

0 0 0

C 9

0.4212962963 0 6.5694444444

T 1

0.3981481481 0 7.3981481481

T 2

0 0.2222222222 0.1111111111 0.1111111111 0.1111111111 0.1111111111 0.1111111111 -6.41164E-16 0.2222222222 0 0.4212962963 0.8425925926 0.3981481481 0.4212962963 0.4212962963 0.4444444444 0.4212962963 0.4212962963 0.4212962963 0.4444444444


0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

0.3981481481 0 6.7314814815

T 3

0 0.1111111111 -6.59544E-16 0.1111111111 -0.111111111 -6.80361E-16 -0.111111111 -7.03875E-16 -7.50593E-16 0 0.4212962963 0.3981481481 0.8425925926 0.4212962963 0.4212962963 0.4444444444 0.4212962963 0.4212962963 0.4212962963 0.4444444444

0.4212962963 0 5.3657407407

T4

0 -8.59604E-16 -7.08048E-16 0.1111111111 -0.111111111 -7.21926E-16 -7.52338E-16 -0.111111111 0.1111111111 0 0.4212962963 0.4212962963 0.4212962963 0.8425925926 0.3981481481 0.4444444444 0.4212962963 0.3981481481 0.4212962963 0.4444444444

224

Intercept R 1 R 2 R 3 R 4 C 1 C 2 C 3 C 4 C 5 C 6 C 7 C 8 C 9 T 1 T 2 T 3 T 4 T 5 T 6 T 7 T 8 T 9 T 10

T 5 -0.37962963 -0.083333333 -0.041666667 -0.041666667 0 0.1111111111 -6.10521E-16 -7.265E-16 -6.70087E-16 -0.111111111 -0.111111111 -6.53118E-16 0.1111111111 0 0.4212962963 0.4212962963 0.4212962963 0.3981481481 0.8425925926 0.4444444444 0.4212962963 0.3981481481 0.4212962963 0.4444444444


-0.555555556 -2.16298E-17 -1.63766E-16 -1.12908E-16 0 0.2222222222 0.1111111111 0.2222222222 -7.08466E-16 -6.35922E-16 0.1111111111 0.1111111111 0.2222222222 0 0.4444444444 0.4444444444 0.4444444444 0.4444444444 0.4444444444 0.8888888889 0.4444444444 0.4444444444 0.4444444444 0.4444444444

T 6 -0.532407407 -0.041666667 0.0416666667 -1.05476E-16 0 0.1111111111 0.1111111111 0.2222222222 0.1111111111 0.1111111111 -7.09929E-16 0.1111111111 0.2222222222 0 0.3981481481 0.4212962963 0.4212962963 0.4212962963 0.4212962963 0.4444444444 0.8425925926 0.4212962963 0.3981481481 0.4444444444

T 7

T 8 -0.490740741 -0.083333333 -0.041666667 -0.041666667 0 0.2222222222 -6.29051E-16 0.2222222222 0.1111111111 0.1111111111 0.1111111111 0.1111111111 0.1111111111 0 0.4212962963 0.4212962963 0.4212962963 0.3981481481 0.3981481481 0.4444444444 0.4212962963 0.8425925926 0.4212962963 0.4444444444


-0.421296296 -0.041666667 0.0416666667 -8.40883E-17 0 0.1111111111 -0.111111111 -7.48079E-16 -0.1111111111 -6.44453E-16 -6.91857E-16 -6.73674E-16 0.1111111111 0 0.3981481481 0.4212962963 0.4212962963 0.4212962963 0.4212962963 0.4444444444 0.3981481481 0.4212962963 0.8425925926 0.4444444444

T 9

-0.444444444 -2.56719E-17 -1.38442E-16 -8.45796E-17 0 0.1111111111 -0.1111111111 0.1111111111 -7.945E-16 -7.4153E-16 -0.111111111 -0.111111111 0.1111111111 0 0.4444444444 0.4444444444 0.4444444444 0.4444444444 0.4444444444 0.4444444444 0.4444444444 0.4444444444 0.4444444444 0.8888888889

T 10

225

T 11 T 12 y

0.4212962963 0 4.4768518519

Intercept R 1 R 2 R 3 R 4 C 1 C 2 C 3 C 4 C 5 C 6 C 7 C 8 C 9 T 1 T 2 T 3 T 4

0.4212962963 0 6.3472222222

0.4212962963 0 5.0324074074

-0.421296296 -0.041666667 -1.28995E-16 0.0416666667 0 -7.20447E-16 -0.111111111 0.1111111111 -6.75091E-16 -0.111111111 -6.70031E-16 -6.46005E-16 0.1111111111 0 0.4212962963 0.3981481481 0.3981481481 0.4212962963

T 11 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

T 12

y

0.4212962963 0 6.4583333333

34.208333333 -25.54166667 -24.16666667 -12.45833333 0 3 1.4444444444 5 1.5555555556 0.7777777778 6.3333333333 2.8888888889 5 0 6.5694444444 7.3981481481 6.7314814815 5.3657407407


0.4444444444 0 8.5555555556

0.4444444444 0 8.4444444444

226

Table 6.3 (Continued ) 5 6 7 8 9 10 11 12

0.4212962963 0.4444444444 0.4212962963 0.4212962963 0.4212962963 0.4444444444 0.8425925926 0 0.6203703704

0 0 0 0 0 0 0 0 0

L1 L2 L3 L4 L1-L2-L3-L4

1 2 3 4 1 2 3 4 5 6 7

Intercept R R R R C C C C C C C

L6 L7 L8 L9 L10 L11 L12

Coefficients

Effect

General Form of Estimable Functions

T T T T T T T T y

4.4768518519 8.5555555556 6.3472222222 5.0324074074 6.4583333333 8.4444444444 0.6203703704 0 338.76851852

227

1 2 3 4 5 6 7 8 9 10 11 12

T T T T T T T T T T T T

18.86789

0.927051

4305.120370 338.768519 4643.888889 Coeff Var

22 13 35

Model Error Corrected Total

Sum of Squares

5.104813

Root MSE

y Mean

7.51

F Value

27.05556

195.687290 26.059117

Mean Square

L15 L16 L17 L18 L19 L20 L21 L22 L23 L24 L25 L1-L15-L16-L17-L18-L19-L20-L21-L22-L23-L24-L25

L13 L1-L6-L7-L8-L9-L10-L11-L12-L13

R-Square

DF

Source

Dependent Variable: y

8 9

C C

0.0003

Pr > F

228


1134.509259 14.046296 16.798822

Mean Square

1317.148148 21.111111 16.798822

Mean Square

3.2997161 3.2997161 3.2997161 3.2997161 3.2997161 3.2997161 3.2997161 3.2997161

Standard Error

Least Squares Means

3403.527778 112.370370 184.787037

28.1250000 28.9537037 28.2870370 26.9212963 26.0324074 30.1111111 27.9027778 26.5879630

3 8 11

R C T

Type III SS

1 2 3 4 5 6 7 8

DF

Source

3951.444444 168.888889 184.787037

y LSMEAN

3 8 11

R C T

Type I SS

T

DF

Source

0. To illustrate the method according to (12.8), however, we develop the ν ∗ (z) in Table 12.6. We obtain ν ∗ (001) = 20 10 = = 3. These 5 d.f. account then for the 5 d.f. among blocks. 2, ν ∗ (101) = 30 10 We note here that the design given in Example 12.3 is slightly better than what could have been obtained with the Kronecker product method of Section 12.2. Both result in confounding main effects (2 d.f. from A3 vs. 1 d.f. from A1 and 2 d.f. from A3 ), but in this situation that seems to be unavoidable. This shows that the GC/n method leads to useful designs not available otherwise. For a detailed discussion of other properties of GC/n designs we refer the reader to Gupta and Mukerjee (1989). 12.4 METHOD OF FINITE RINGS The methods of constructing systems of confounding for symmetrical factorials as discussed in Chapters 8, 9, 10, and 11 (Sections 11.3, 11.7, and 11.8) are based on the arithmetic in Galois fields. To transfer these procedures to asymmetrical

481

METHOD OF FINITE RINGS z

Table 12.6 Determination of ρiz11 ρiz22 ρi33 for the Design of Table 12.5 z i1

i2

i3

100

010

110

001

101

011

111

1 2 1 2 1 2 1 2 1 2

1 1 2 2 3 3 4 4 5 5

1 4 1 4 1 4 1 4 1 4

1 −1 1 −1 1 −1 1 −1 1 −1

4 4 −1 −1 −1 −1 −1 −1 −1 −1

4 −4 −1 1 −1 1 −1 1 −1 1

5 −1 5 −1 5 −1 5 −1 5 −1

5 1 5 1 5 1 5 1 5 1

20 −4 −5 1 −5 1 −5 1 −5 1

20 4 −5 −1 −5 −1 −5 −1 −5 −1

0

0

0

20

30

0

0

Sum

factorials involves combining elements from different finite fields. Using results from abstract algebra (e.g., van der Waerden, 1966), White and Hultquist (1965) and Raktoe (1969, 1970) provided the foundation for generalizing the methods for symmetrical factorials to asymmetrical factorials (Hinkelmann, 1997). We shall give a brief description of the method (following Raktoe, 1969) without going into all mathematical details and proofs. 12.4.1 Mathematics of Ideals and Rings First, we need the following mathematical results. Suppose we have m distinct primes p1 , p2 , . . ., pm . Let GF(pi ) be the Galois field associated withpi (i = 1, 2, . . . , m), the elements of which are the residues mod pi . Let q = m i=1 pi and R(q) be the ring of residue classes mod q. We denote by I (w) an ideal generated by an arbitrary element w of R(q). We then state the following results (for proofs see Raktoe, 1969): 1. The element aj =

m i=1 i=j

pi − pj = c j − pj

with cj =

pi

i=j

in R(q) is prime to q and hence aj−1 exists. The aj ’s belong to the multiplicative group of nonzero divisors in R(q). 2. The element bj = cj aj−1 generates the ideal I (bj ) in R(q). 3. The element bj is the multiplicative identity in I (bj ). 4. The multiplicative identity element 1 in R(q) is the sum of the multiplicative identities in I (bj ), that is, 1 = m j =1 bj .

482

CONFOUNDING IN ASYMMETRICAL FACTORIAL DESIGNS

5. The ring R(q) is the direct sum of ideals I (bj ), that is, R(q) = m j =1 ⊗I (bj ). 6. The field GF(pj ) is isomorphic to the ideal I (bj ); for x ∈ GF(pj ) and y ∈ I (bj ) the mapping σ : GF(pj ) → I (bj ) is defined by σ (x) = bj x = y. 7. Addition and multiplication of elements from different Galois fields x ∈ GF(pj ), x ∗ ∈ GF (pi ) are defined by x + x ∗ = σ (x) + σ (x ∗ ) mod q xx ∗ = σ (x)σ (x ∗ ) mod q 8. Addition and multiplication of an element r ∈ R(q) and an element x ∈ GF (pj ) are defined by r + x = r + σ (x) mod q rx = rσ (x) mod q 9. The ring R(q) is the direct sum of the GF(pj ), that is, R(q) = GF (pj ).

m

j =1 ⊗

To illustrate these results, we consider the following example. Example 12.4 Suppose we have p1 = 2, p2 = 3. Then p = 6 and the generators of the ideals are b1 = 3(3 − 2)−1 = 3 · 1 = 3 mod 6 and

b2 = 2(2 − 3)−1 = 2(5)−1 = 2 · 5 = 4 mod 6

[(5)−1 = 5 since 5 · 5 = 25 = 1 mod 6]. We thus have the two mappings GF(2) → I (3) 0 1

σ − →

and

0 3

GF(3) → I (4)   0   0 τ 4 1 − →   2 2

and the direct sums GF(2) ⊗ GF(3)

=

I (3) ⊗ I (4)

=

R(6)

0+0 1+0 0+1 1+1 0+2 1+2

= = = = = =

σ (0) + τ (0) σ (1) + τ (0) σ (0) + τ (1) σ (1) + τ (1) σ (0) + τ (2) σ (1) + τ (2)

= = = = = =

0 3 4 1 2 5

483

METHOD OF FINITE RINGS

12.4.2 Treatment Representations We now apply these concepts and results to asymmetrical factorials of the form nm . For ease of notation we shall consider specifically the p1n1 × p2n2 × · · · × pm 2 2 2 × 3 factorial. Generalizations will then become obvious. Ordinarily, treatment combinations would be represented in the form of quadruples, say x = (x11 , x12 , x21 , x22 ), where x11 , x12 refer to the levels of the two-level factors, A1 , A2 say, and x21 , x22 refer to the levels of the threelevel factors, B1 , B2 , say, that is, x1i = 0, 1, x2i = 0, 1, 2. We shall refer to this as the elementary representation of treatment combinations and denote it by T (x), remembering that the components in x are elements from the respective Galois fields. In order to perform arithmetic with treatment combinations, we now replace T (x) by a new representation T (v), where the components in v = (v11 , v12 , v21 , v22 ) represent the corresponding elements from the respective ideals, that is, v11 , v12 ∈ I (3), v21 , v22 ∈ I (4). The correspondence is as follows: T (x)

T (v)

x11

x12

x21

x22

0 0 1 1 0 0 0

0 1 0 1 0 0 0

0 0 0 0 0 0 1

0 0 0 0 1 2 0

v11 0 0 3 3 0 0 0

v12

v21

v22

0 3 0 3 0 0 0

0 0 0 0 0 0 4

0 0 0 0 4 2 0

2

2

.. . 1

1

.. . 2

2

3

3

12.4.3 Representation of Main Effects and Interactions Concerning the representation of main effects and interactions, we recall that for a p n factorial an effect E α is represented by contrasts among the “levels” α of E α . Here the components in α = (α1 , α2 , . . ., αn ) also E0α , E1α , . . ., Ep−1 take on the values 0, 1, . . ., p − 1 with the understanding that the first nonzero αi equals 1. We note also that for the p n factorial we can write an interaction in general as Az11 × Az22 × · · · × Aznn , where the zi are either 0 or 1 with the understanding that if zi = 0 then the letter Ai is dropped from the interaction. An interaction Ai1 × Ai2 × · · · × Ai consists then of (p − 1)−1 components Ai1 Aαi22 · · · Aαi where the αi take on all possible values between 1 and p − 1. Each such component represents p − 1 d.f. To generalize this representation for asymmetrical factorials, we now write an z∗ z∗ z∗ z∗ ∗ and interaction for the 22 × 32 factorial as A111 × A212 × B1 21 × B2 22 , where z11

484


∗ are either 0 or 3, that is, elements in I (3), and z∗ and z∗ are either 0 or 4, z12 21 22 ∗ = 0 then that factor is that is, elements in I (4), with the understanding that if zij dropped from the interaction. We may have, for example, A31 × B14 × B24 , which refers to the interaction among the factors A1 , B1 , and B2 . The components of this interaction are, in analogy to the symmetrical component systems, given by A31 × B14 × B24 = {A31 B14 B24 , A31 B14 B22 }. The number of “levels,” L say, for these interaction components is determined by the fact that they involve factors with two and three levels. The L is equal to the number of residue classes for the direct sum of the ideals corresponding to the factors involved. In our case this is the number of residue classes for I (3) ⊗ I (4) = R(6), that is, L = 6. More generally, if the interaction involves only factors out of m, say Fi1 , Fi2 , . . ., Fi ( < m), then L is equal to the number of residue classes in I (bi1 ) ⊗ I (bi2 ) ⊗ · · · ⊗ I (bi ) that is generated by the greatest common divisor of bi1 , bi2 , . . ., bi . Trivially, this includes the main effects, that is, for main effect Fi with pi levels, L equals the number of residue classes for I (bi ), which equals of course pi . The reason why we mention the levels of an interaction is that the total number of treatment combinations can be partitioned into L equal-sized sets (of treatment combinations) that can be identified with blocks and a recognizable system of confounding. Let us consider the 22 × 32 factorial. For it we have for the various main effects and interactions the following L’s:

L(A31 ) = L(A32 ) = L(A31 A32 ) = 2 L(B14 ) = L(B24 ) = L(B14 B24 ) = L(B14 B22 ) = 3 ϕ

ϕ

η

η

L(A31 B14 ) = L(A31 B24 ) = · · · = L(A1 1 A1 2 B1 1 B2 2 ) = · · · = L(A31 A32 B14 B22 ) = 6 where ϕ1 , ϕ2 , ∈ I (3), η1 , η2 ∈ I (4). For A31 A32 B14 B22 , for example, this leads to the partitioning of the 36 treatment combinations (v11 , v12 , v21 , v22 ) into six sets according to the equations. 3v11 + 3v12 + 4v21 + 2v22 = δ

(12.9)

where δ = 0, 1, 2, 3, 4, 5, that is, δ ∈ R(6). The sets are given in Table 12.7. Each set represents a block of size 6. We note that the set δ = 0 represents the IBSG and all other sets can be obtained easily by adding to each treatment combination in the IBSG a treatment not in the IBSG and not in any other already constructed set. As a consequence of constructing blocks according to (12.9), it follows that the interaction associated with (12.9), that is, A31 A32 B14 B22 , is confounded with blocks. This statement, however, is to be understood only in the sense that together with A31 A32 B14 B22 also A31 A32 and B14 B22 , that is, the components from the symmetrical systems, are confounded with blocks. This is so because to satisfy (12.9) for a given δ we must have 3v11 + 3v12 = δ1 mod 6

(12.10)

485


Table 12.7 Sets of Treatment Combinations Satisfying 3v11 + 3v12 + 4v21 + 2v22 = δ δ=0 0 3 0 3 0 3

0 3 0 3 0 3

0 0 4 4 2 2

δ=1 0 0 4 4 2 2

δ=3 3 0 3 0 3 0

0 3 0 3 0 3

0 0 4 4 2 2

3 0 3 0 3 0

0 3 0 3 0 3

4 4 2 2 0 0

δ=2

0 0 4 4 2 2

0 3 0 3 0 3

δ=4 0 0 4 4 2 2

0 3 0 3 0 3

0 3 0 3 0 3

4 4 2 2 0 0

0 3 0 3 0 3

2 2 0 0 4 4

0 0 4 4 2 2

δ=5

0 0 4 4 2 2

3 0 3 0 3 0

0 3 0 3 0 3

2 2 0 0 4 4

0 0 4 4 2 2

and 4v21 + 2v22 = δ2 mod 6

(12.11)

with δ1 + δ2 = δ. Thus we see that the system of confounding generated in this manner is the same as that obtained by using the Kronecker product method of Section 12.2 by considering confounding in the 22 and 32 systems separately and then combining them. This is true in general (see Voss, 1986) and thus may detract somewhat from the value of this method, but we shall discuss below that it provides for an extension of a useful result for symmetrical factorials that other methods do not do. 12.4.4 Parameterization of Treatment Responses Recall from Chapter 11 that we expressed the true response a(x) of a treatment combination x as a linear function of interaction components E α as follows: a(x) = M +

Eαα x

(12.12)

α

where α is the summation over all partitions α. To write down the counterpart to (12.12) for asymmetrical factorials, we consider again the 22 × 32 factorial. The partitions α are determined by the various main effects and interactions and are given in Table 12.8. Associated with each partition is an effect/interaction component and each component has a number, L, of levels as mentioned earlier.

486


Table 12.8 Partitions for the 22 × 32 Factorial Interaction/Modified Interaction Component

Interaction ∗ , z∗ , z∗ , z∗ ) z∗ = (z11 12 21 22

Partition α = (α11 , α12 , α21 , α22 )

3000

3000

(A31 )δ (δ = 0, 3)

0300

0300

(A32 )δ (δ = 0, 3)

3300

3300

(A31 A32 )δ (δ = 0, 3)

0040

0040

(B14 )δ (δ = 0, 4, 2)

0004

0004

(B24 )δ (δ = 0, 4, 2)

0044

0044

(B14 B24 )δ (δ = 0, 4, 2)

0042

(B14 B22 )δ (δ = 0, 4, 2)

3040

(A31 B14 )∗δ = (A31 B14 )δ −

3040

(A31 )δ1 − (B14 )δ2 (δ = 0, 1, 2, 3, 4, 5; δ1 = 0, 3; δ2 = 0, 4, 2) 3004

3004

(A31 B24 )∗δ = (A31 B24 )δ − (A31 )δ1 − (B24 )δ2

0340

0340

(A32 B14 )∗δ = (A32 B14 )δ − (A32 )δ1 − (B14 )δ2

0304

0304

(A32 B24 )∗δ = (A32 B24 )δ − (A32 )δ1 − (B24 )δ2

3340

3340

(A31 A32 B14 )∗δ = (A31 A32 B14 )δ − (A31 A32 )δ1 − (B14 )δ2

3304

3304

(A31 A32 B24 )∗δ = (A31 A32 B24 )δ − (A31 A32 )δ1 − (B24 )δ2

3044

3044

(A31 B14 B24 )∗δ = (A31 B14 B24 )δ − (A31 )δ1 − (B14 B24 )δ2

3042

(A31 B14 B22 )∗δ = (A31 B14 B22 )δ − (A31 )δ1 − (B14 B22 )δ2

487


Table 12.8 (Continued ) Interaction ∗ , z∗ , z∗ , z∗ ) z∗ = (z11 12 21 22

Partition α = (α11 , α12 , α21 , α22 )

0344

0344

Interaction/Modified Interaction Component (A32 B14 B24 )∗δ = (A32 B14 B24 )δ − (A32 )δ1 − (B14 B24 )δ2

0342

(A32 B14 B22 )∗δ = (A32 B14 B22 )δ − (A32 )δ1 − (B14 B22 )δ2

3344

3344

(A31 A32 B14 B24 )∗δ = (A31 A32 B14 B24 )δ − (A31 A32 )δ1 − (B14 B24 )δ2

3342

(A31 A32 B14 B22 )∗δ = (A31 A32 B14 B22 )δ − (A31 A32 )δ1 − (B14 B22 )δ2

The levels are the elements δ in I (3), I (4), R(6), respectively, as indicated in Table 12.8. We define the components as Eδα = (Aα1 11 Aα2 12 B1α21 B2α22 )δ = {average of all treatment responses a(v11 , v12 , v21 , v22 ) with α11 v11 + α12 v12 + α21 v21 + α22 v22 = δ(mod q)} −{average of all treatment responses}

(12.13)

where in our case q = 6. For interactions involving factors with different numbers of levels we now introduce what we shall call modified interaction components defined as (E α )∗δ = (Aα1 11 Aα2 12 B1α21 B2α22 )∗δ = (Aα1 11 Aα2 12 B1α21 B2α22 )δ −(Aα1 11 Aα2 12 )δ1 − (B1α21 B2α22 )δ2

(12.14)

We shall refer to the partitions α involving only factors with the same number of levels as pure partitions and those that involve factors not all having the same

488


number of levels as mixed partitions. Let A1 denote the set of all pure partitions and A2 the set of all mixed partitions. We can then write Eαα v + (E α )∗α v (12.15) a(v) = M + α∈A1 α∈A2 where M represents the overall mean. Expression (12.15) is the extension of (12.12) for the asymmetrical factorial and is used in the same way as (12.12) is used for the symmetrical factorial. 12.4.5 Characterization and Properties of the Parameterization The proposed parameterization (12.15) is a major result in that it unifies the theory of symmetrical and asymmetrical factorials, both with respect to the construction of design, that is, systems of confounding, and the analysis of such designs. Although we have only illustrated the idea in terms of a simple example, the result is true in general. The same holds for the following comments about further characterizations and properties of the modified components [defined in (12.14)] as part of the parameterization (12.15). 1. Expressing each a(v) as (12.15) represents a reparameterization of the treatment This reparameterization is singular since for each E α we have effects. α δ Eδ = 0. 2. For the modified components the values for δ1 and δ2 are determined uniquely from the addition table, for example, δ2

δ1

0

4

2

0

0

4

2

3

3

1

5

where the entries in the table are δ = δ1 + δ2 . 3. For the modified components we have δ1 (E α )∗δ1 +δ2 = 0 for each δ2 and α ∗ δ2 (E )δ1 +δ2 = 0 for each δ1 . 4. The degrees of freedom ν(α) associated with the component E α are given by ν(α) = L(E α ) − 1; for example, for A31 A32 B14 B22 we have L(A31 A32 B14 B22 ) = 6 and hence ν(3, 3, 4, 2) = 5. 5. The degrees of freedom ν ∗ (α) associated with the modified component (Aα1 11 Aα2 12 B1α21 B2α22 )∗ are given by ν ∗ (α11 α12 α21 α22 ) = ν(α11 α12 α21 α22 ) − ν(α11 α12 00) − ν(00α21 α22 ) = ν(α11 α12 00) · ν(00α21 α22 )

489


6. If a block design is constructed by confounding Aα1 11 Aα2 12 B1α21 B2α22 and if Aα1 11 Aα2 12 B1α21 B2α22 is a mixed component, it follows from comment 3 and the definition of the modified component that Aα1 11 Aα2 12 and B1α21 B2α22 are also confounded (as pointed out earlier). 7. The component Eδα is estimated by α = {mean of observed responses for treatments v satisfying E δ α v = δ} − {overall observed mean} 8. If the confounded interactions, that is, Aα1 11 Aα2 12 B1α21 B2α22 , Aα1 11 Aα2 12 , B1α21 B2α22 , are assumed to be negligible, then differences between two treatment responses (more generally, linear contrasts among treatment responses), say a(v) − a(w), can be estimated by a (v) − (w) =

α − E α E αv αw α∈A1 α )∗ α )∗ − (E (E αv αw α∈A2

+

(12.16)

where refers to summation over all α except those belonging to confounded interactions. Using (12.14), we find, for example, a(3040) − a(0000) = (A31 )3 − (A31 )0 + (A31 A32 )3 − (A31 A32 )0 + (B14 )4 − (B14 )0 + (B14 B24 )4 − (B14 B24 )0 + (A31 B14 )∗1 − (A31 B14 )∗0 + (A31 B24 )∗3 − (A31 B24 )∗0 + (A32 B14 )∗4 − (A32 B14 )∗0 + (A31 A32 B14 )∗1 − (A31 A32 B14 )∗0 + (A31 A32 B24 )∗3 − (A31 A32 B24 )∗0 + (A31 B14 B24 )∗1 − (A31 B14 B24 )∗0 + (A31 B14 B22 )∗1 − (A31 B14 B22 )∗0 + (A32 B14 B24 )∗4 − (A32 B14 B24 )∗0

490


+ (A32 B14 B22 )∗4 − (A32 B14 B22 )∗0 + (A31 A32 B14 B24 )∗1 − (A31 A32 B14 B24 )∗0

(12.17)

Observed means (based on different numbers of observations) are then substituted for the true means to obtain a (3040) − a (0000). 3 )3 − (A 3 )0 can 9. It is obvious that variances for differences of the form (A 1 1 be written out easily in terms of the number of observations satisfying the equation 3v11 = δ for a given δ. If N is the total number of treatment combinations, then the number of treatment combinations satisfying 3v11 = δ is given by N/L(A31 ). Hence L(A31 ) 2 3 )3 − (A 3 )0 = 2 var (A σe 1 1 N which equals 19 σe2 for the 22 × 32 factorial. 3 B 4 )∗ − (A 3 B 4 )∗ , that is, involving modFor differences of the form (A 1 2 3

1 2 0

ified components, it can be worked out that 3 B 4 )∗ − (A 3 B 4 )∗ = 2 L(A3 B 4 ) − L(A3 ) − L(B 4 ) σ 2 var (A e 2 1 2 1 1 2 3 1 2 0 N which for the 22 × 32 factorial equals 1 2 2 [6 − 2 − 3]σe2 = σ 36 18 e Finally, it can be shown that the covariances between differences of components are zero, for example, 4 B 2 ) − (B 4B 2) 3 )0 , (B 3 )3 − (A =0 cov (A 4 0 1 1 1 2 1 2 cov cov

3 B 4 )∗ − (A 3 B 4 )∗ 3 )0 , (A 3 )3 − (A =0 (A 1 1 1 1 1 1 1 0

3 B 4 )∗ − (A3 B 4 )∗ , (A 3 A3 B 4 )∗ − (A 3 A3 B 4 )∗ =0 (A 1 1 0 1 1 1 1 2 1 1 1 2 1 0

10. Using (12.16) and the results in comment 9, one can obtain easily the variances of the estimates of treatment contrasts, that is, var[ a (v) − a (w)] = var[ τ (v) − τ (w)] = (sum of the variances of all estimated component differences)

491

BALANCED FACTORIAL DESIGNS (BFD)

for example, for (12.17) we find var[ a (3040) − a (0000)] =

2 2 2 2 + + + 18 18 12 12 2 2 2 20 2 − − σ + 10 σe2 = 6 18 12 18 e

11. The sums of squares for unconfounded effects can be written out similarly to those for symmetrical factorials: SS(A31 ) =

2 N 3 )δ (A 1 L(A1 ) δ∈I (3)

SS(B14 B22 ) = SS(A31 B14 ) =

2 N 4B 2) ( B δ 1 2 L(B14 B22 ) δ∈I (4) N

L(A31 B14 ) δ∈R(6)

3 B 4 )∗ (A 1 1 δ

2

and because of comment 9 these sums of squares are orthogonal to each other. This makes testing of hypotheses concerning main effects and interactions easy. 12.4.6 Other Methods for Constructing Systems of Confounding In addition to the method discussed in this section, alternative but essentially equivalent methods of constructing systems of confounding for asymmetrical factorials have been proposed. A method based on finite groups and rings is discussed by Banerjee (1970), Worthley and Banerjee (1974) (see also Raktoe, Rayner, and Chalton, 1978), and Sihota and Banerjee (1981), who extended the method to include any number of levels, thus avoiding the use of pseudofactors. A different technique based on the Chinese remainder theorem was introduced by Lin (1986) and extended by Huang (1989).

12.5 BALANCED FACTORIAL DESIGNS (BFD) The incomplete block designs for asymmetrical factorials we have discussed so far have resulted in the confounding of certain contrasts belonging to some main effects and/or interactions. As a consequence the number of degrees of freedom for such effects has been reduced (possibly to zero). The lost degrees of freedom are said to be confounded with blocks. As we have seen in the previous sections, the confounding of interactions may sometimes also lead to the confounding of main effects, for example, with the methods of Sections 12.2 and 12.4. If we

492


confound, for example, a two-factor interaction involving factors with different numbers of levels, then the corresponding main effects are confounded also. This is generally an undesirable feature. If it cannot be avoided, the method of partial confounding, that is, using different systems of confounding in different replicates, provides some relief. This, however, may not always be feasible from a purely practical point of view because of the usually large number of experimental units required. In this section we shall discuss a method that results in a different kind of partial confounding. Rather than confound completely certain degrees of freedom in some replicates and not at all in others, the method to be described results in some loss of information for some degrees of freedom over the whole design. We shall illustrate this with a very simple example. Example 12.5 Consider the 22 × 3 factorial in blocks of size 6. Using the Kronecker product method of Section 12.2, a reasonable design may be obtained by using three replicates of two blocks each. If we denote the factors by A1 , A2 , and B with two, two, and three levels, respectively, we may construct the design by confounding A1 in replicate I, A2 in replicate II, and A1 A2 in replicate III. The allocation of the treatment combinations is given in Table 12.9a. Compared to a RCBD with b = 3 blocks (replicates), the relative information for main effects A1 , A2 and the interaction A1 A2 is 23 and 1 for all other effects. In Table 12.9b an arrangement due to Yates (1937b) is given that is the only arrangement not resulting in the confounding of main effects. Kempthorne (1952) Table 12.9 22 × 3 Factorial in Blocks of Size 6 Replicate I Block

1

II 2

3

III 4

5

6

010 011 012 110 111 112

000 001 002 110 111 112

100 101 102 010 011 012

000 110 101 011 002 112

100 010 101 011 002 112

000 110 001 111 102 012

a. Kronecker Product Method 000 001 002 010 011 012

100 101 102 110 111 112

000 001 002 100 101 102 b. Yates Method

000 110 101 011 102 012

100 010 001 111 002 112

100 010 001 111 102 012

493


shows, from first principles, using the least-squares method, that the relative information for A1 A2 is 89 , for A1 A2 B is 59 , and for all other effects is 1. If we define the total relative loss in information, TRL say, as the weighted sum of relative losses for the various effects, with the weights being the degrees of freedom for each effect, we find that TRL is the same for both arrangements, namely TRL(a) = 1 ·

1 3

+1·

1 3

+1·

TRL(b) = 1 ·

1 9

+2·

4 9

=1

1 3

=1

Even so, we would generally prefer arrangement (b) over arrangement (a) since its loss of information is mainly associated with the three-factor interaction. Both designs of Example 12.5 belong to the class of balanced factorial designs (BFD), a notion defined by Shah (1958) for the general factorial experiment. Although Shah (1958, 1960) pointed out the connection between BFDs and PBIB designs, Kurkjian and Zelen (1963) and Kshirsagar (1966) established the oneto-one correspondence between BFDs and EGD-PBIB designs as defined by Hinkelmann and Kempthorne (1963) and Hinkelmann (1964) (see Section 4.6.9). We shall now give some definitions of and results for BFDs and then give some methods of construction. 12.5.1 Definitions and Properties of BFDs Suppose we have n factors A1 , A2 , . . ., An with s1 , s2 , . . ., sn levels, respectively, where the si (i = 1, 2, . . . , n) do not necessarily have to be all distinct. Following Kurkjian and Zelen (1963) and extending the notation of Section 11.9, we define the si × si matrix M i = si I − II

(12.18)

and M zi i

=

Isi

if zi = 0

Mi

if zi = 1

M z = M z11 × M z22 × · · · × M znn

(12.19)

for z = (z1 , z2 , . . . , zn ) and zi = 0, 1(i = 1, 2, . . ., n). Let Z = {z} be the collection of all possible vectors z except z = (0, 0, . . ., 0). Further, let t = ni=1 si denote the number of treatment combinations. Then for any z ∈ Z the interaction components in the n-factor factorial can be expressed in terms of the treatment effects τ (x1 , x2 , . . ., xn )(0 ≤ xi ≤ si − 1; i = 1, 2, . . ., n) arranged in standard order (see Section 11.9) in the vector τ as az =

1 z M τ t

(12.20)

494


Let φ(z) = z a z = ϕ (z)τ

(12.21)

denote a normalized contrast belonging to a z (note that we are using a z for two different things but it should always be clear from the context what is meant). With τ = C−Q as a solution to the reduced normal equations we obtain the BLUE of φ(z) in (12.21) as 1 a z = z M z τ φ(z) = z t

(12.22)

Following Gupta and Mukerjee (1989), we then give the following definition. Definition 12.1 In a factorial design an interaction az (z ∈ Z) is said to be balanced if either (a) all contrasts belonging to a z are estimable and the BLUEs of all normalized contrasts belonging to a z have the same variance or (b) no contrast belonging to a z is estimable. If we adopt the convention that if a contrast is nonestimable its BLUE has variance ∞, then parts (a) and (b) in Definition 12.1 can be combined, and it is in this sense that we shall use the definition. For an overall characterization of a factorial design, we then have the following definition. Definition 12.2 A factorial design is said to be a BFD if the interaction a z is balanced for every z ∈ Z. It is of historical interest to note that Li (1944, p. 457) already refers to BFD (although not by that name) when he says, “An interaction which has more than 1 degree of freedom should be confounded as evenly as possible.” Gupta and Mukerjee (1989) show that an interaction a z is balanced in the sense of Definition 12.1 if and only if the BLUEs of any two orthogonal contrasts belonging to a z are uncorrelated. Recall that a similar result holds for factorial designs with OFS except that in that case BLUEs of contrasts belonging to different interactions are uncorrelated. Hence in a BFD with OFS we have between-interaction and within-interaction orthogonality (Gupta and Mukerjee, 1989). As we have pointed out earlier (e.g., I.7), orthogonality is a useful property as it leads to ease of analysis and ease of interpretation of the results from an experiment. We have seen in Chapter 11 that OFS requires a certain structure of C − (see Theorem 11.12) and hence of C and ultimately N N . The results of Shah (1958, 1960), Kurkjian and Zelen (1963), Kshirsagar (1966), and Gupta and Mukerjee (1989) have led to similar structures for NN (and hence C and

495


C − ) for BFDs. We shall give a brief description of the major results, deferring details to the above references. The basic result for giving an algebraic characterization of a BFD with OFS is what Kurkjian and Zelen (1963) have called property A of the matrix N N . Let D δi i be an si × si matrix defined as D δi i

=

I si

if δi = 0

Isi I si

if δi = 1

for i = 1, 2, . . . , n, and Dδ =

n

×D δi i

i=1

with δ = (δ1 , δ2 , . . . , δn ). Denote the set of all δ (with δi = 0, 1) by Z ∗ = Z ∪{(0, 0, . . . , 0)}. We then state formally: nDefinition 12.3 (Gupta and Mukerjee, 1989) A t × t matrix G with t = i=1 si is said to have property A if it is of the form G=

h(δ)D δ

(12.23)

δ Z ∗

where the h(δ) are constants depending on δ. Now suppose the matrix C has property A, that is, 1 C = rI − NN = h(δ)D δ k ∗

(12.24)

δ Z

We want to show that if C is of the form (12.24) then all normalized contrasts belonging to a z are estimated with the same variance and the estimators of any two orthogonal contrasts belonging to a z are uncorrelated. Let H i be an (si − 1) × si matrix of normalized orthogonal contrasts, that is, H i Isi = 0 and H i H i = I si −1 , and define H zi i

=

Hi

if zi = 1

1

if zi = 0

and Hz =

n i=1

×H zi i

496


A complete set of normalized orthogonal contrasts belonging to a z is then given by H z a z . More specifically, the contrasts belonging to the p-factor interaction Ai1 × Ai2 × · · · × Aip (1 ≤ p ≤ n), that is, to the si1 si2 · · · sip elements of ai1 ⊗ ai2 ⊗ · · · ⊗ aip , are given by

H i1 × H i2 × · · · × H ip

ai1 ⊗ ai2 ⊗ · · · ⊗ aip

!

and there are (si1 − 1)(si2 − 1) · · · (sip − 1) such contrasts. We can then prove the following theorem. Theorem 12.1 If for an s1 × s2 × · · · × sn factorial design the C matrix satisfies property A, then the design is a BF D and the variance–covariance matrix for H z a z is given by az) = var(H z

1 1−zi rE(z) i si

I σe2

(12.25)

where r is the number of replications for each treatment combination and E(z), the efficiency factor for a z , is a constant depending only on z. Proof

Let C=

h(δ)D δ

δ Z ∗

Then M zC =

h(δ)M z D δ

δ Z ∗

=

δ Z ∗

Now and

h(δ)

×M zi i D δi i

(12.26)

i

M zi i D δi i = (1 − zi δi )si(1−zi )δi M zi i

×M zi i D 1i = 0

i

so that (12.26) can be written as M zC =

h(δ)

δ∈Z ∗∗

= rE(z)M

(1−z)δi × (1 − zi δi )si M zi i

i z

(12.27)

497


where, by definition,

rE(z) =

h(δ)

δ∈Z ∗∗

(1 − zi δi )si(1−zi )δi

(12.28)

i

and Z ∗∗ = Z ∗ − {(1, 1, . . . , 1)}. It follows now from C τ =Q and (12.27) that τ = M zQ M z C can be written as τ = M zQ rE(z)M z and, using (12.20), rtE(z) az = M zQ so that az = H z

1 H zM zQ rtE(z)

Then var(H z az) =

1 H z M z C(M z ) (H z ) σe2 [rtE(z)]2

and using (12.27) again az) = var(H z

1 H z M z (M z ) (H z ) σe2 t 2 rE(z)

(12.29)

To simplify (12.29), we write H z M z = H z11 × H z22 × · · · × H znn

!

M z11 × M z22 × · · · × M znn

! ! ! = H z11 M z11 × H z22 M z22 × · · · × H znn M znn From the definitions of H zi i and M zi i it follows that H z1i M z1i

=

I si

if zi = 0

si H i

if zi = 1

!

498


and H z M z (M z ) (H z ) = (M z H z )(M z H z ) =

si1−zi

i

[sizi ]2 I (12.30)

i

Substituting (12.30) into (12.29) yields az) = var(H z

1+z 1 si i I σe2 t 2 rE(z) i

1 zi 2 = si I σe trE(z) i

= using the fact that t =

i si .

1 1−zi rE(z) i si

I σe2

This is the desired result.

We comment briefly on the efficiency factors E(z) = (z1 , z2 , . . ., zn ) associated with the interaction a z . Recall that the efficiency factor of an IBD is defined as the ratio of the average variance for simple treatment comparisons using an RCBD over the corresponding variance for the IBD. We saw in Chapter 4 that if the IBD is a PBIB(m) design then we may have up to m different efficiencies E1 , E2 , . . ., Em , where Ei refers to the efficiency factor for comparing two treatments that are ith associates (i = 1, 2, . . ., m). In the context of the BFD we have up to 2n − 1 different efficiency factors E(z) associated with treatment comparisons belonging to the various interactions (which in our terminology also include the main effects). Let z,i a z (i = 1, 2, . . ., ν(z)) be a set of linearly independent normalized contrasts belonging to a z [e.g., the rows of H z constitute such a set of z,i , where ν(z) represents the number of degrees of freedom for a z ]. We know that with an RCBD each such contrast is estimated in terms of treatment means, averaging over the levels of those factors not involved in the interaction, that is, the factors for which zi = 0 in a z . With r replications (blocks) we then have varRCBD ( z,i a z ) =

r

1 1−zi i si

σe2

(12.31)

From (12.25) we have varIBD ( z,i a z ) =

r

1 σe2 1−zi s E(z) i i

(12.32)

The ratio of (12.31) to (12.32) is then equal to E(z). We shall return to evaluating E(z) shortly.

499


We have shown in Theorem 12.1 that if for a factorial design the C matrix has a certain structure, as defined by property A, then the design is a BFD. Kshirsagar (1966) proved the converse so that we can now state the following theorem. Theorem 12.2 For an s1 × s2 × · · · × sn factorial design to be a BFD with OFS, it is necessary and sufficient that the C matrix of the design satisfies property A. We proved the sufficiency part in Theorem 12.1 and for the necessity part we refer the reader to Kshirsagar (1966) (see also Gupta and Mukerjee, 1989). 12.5.2 EGD-PBIBs and BFDs To provide further insight into BFDs, we recall that the association matrices B γ for an EGD-PBIB design can be expressed as Kronecker products of (the trivial) association matrices for a BIBD, B 0 = I and B 1 = II − I (see Section 5.4). More specifically, using the notation of this section, we define I si for γi = 0 γi Bi = Isi I si − I si for γi = 1 Then the association matrix, B γ , for γ th associates [with γ = (γ1 , γ2 , . . ., γn ) and γi = 0, 1 for i = 1, 2, . . ., n] can be written as γ

γ

γ

B γ = B 11 × B 22 × · · · × B nn with NN =

λ(γ )B γ

γ ∈Z ∗

or C=

c(γ )B γ

(12.33)

γ Z ∗

where c(0, 0, . . . , 0) = λ(0, 0, . . . , 0)

k−1 k

1 c(γ ) = − λ(γ ) for all γ = (0, 0, . . . , 0) k It is not difficult to speculate that there exists some relationship between BFDs and EGD-PBIB designs. Indeed, Kshirsagar (1966) proved the following (for an alternative proof see Gupta and Mukerjee, 1989). Theorem 12.3 Every n-factor BFD is an EGD-PBIB (2n − 1) design. This result is useful for several reasons, two of which we shall discuss in some detail:

500


i. For any given EGD-PBIB it is easy to obtain the efficiency factors; ii. Existing methods for constructing EGD-PBIBs can be used to obtain BFDs. Let us consider first i by returning to Example 12.5. Example 12.5 (Continued) For arrangement b in Table 12.9 it is easy to verify, using the association scheme for EGD-PBIBs, that λ(000) = 3 λ(100) = 0 λ(010) = 0 λ(110) = 3

λ(001) = 1 λ(101) = 2 λ(011) = 2 λ(111) = 1

Further, we can express each B γ easily in terms of the D δi i as used in defining property A [see (12.23)]. We note that B 01 = B 02 = I 2 = D 01 = D 02 B 11 = B 12 = I2 I 2 − I 2 = D 11 − D 01 = D 12 − D 02 and

B 03 = I 3 = D 03 , B 13 = I3 I 3 − I 3 = D 13 − D 03

so that, for example, B (000) = D 01 × D 02 × D 03 B (100) = D 11 × D 02 × D 03 − D 01 × D 02 × D 03 B (110) = D 11 × D 12 × D 03 − D 11 × D 02 × D 03 − D 01 × D 12 × D 03 + D 01 × D 02 × D 03 and so on. Collecting terms, we can rewrite (12.33) as C = 3 − 86 D (000) − 46 D (100) − 46 D (010) − 26 D (110) + 26 D (001) − 16 D (101) − 16 D (011) − 16 D (111) that is, in the form (12.23) with h(000) =

10 6

h(100) = − 46

h(010) = − 46

h(110) = − 26

h(001) =

2 6

h(101) = − 16

h(011) = − 16

h(111) = − 16

501


Using (12.28), we can now obtain the efficiency factors E(z). Let us write Wi (zi , δi ) = (1 − zi δi )si(1−zi )δi with   1  1 Wi (zi , δi ) = si    0

for for for for

zi zi zi zi

= 0, = 1, = 0, = 1,

δi δi δi δi

=0 =0 =1 =1

Then (12.28) can be rewritten as

rE(z1 , z2 , . . . , zn ) =

h(z1 , z2 , . . . , zn )

δ∈Z ∗∗

n

Wi (zi , δi )

(12.34)

i=1

So, for example, 6 · 3E(110) = 10W1 (1, 0) W2 (1, 0) W3 (0, 0) −4W1 (1, 1) W2 (1, 0) W3 (0, 0) −4W1 (1, 0) W2 (1, 1) W3 (0, 0) −2W1 (1, 1) W2 (1, 1) W3 (0, 0) +2W1 (1, 0) W2 (1, 0) W3 (0, 1) −W1 (1, 1) W2 (1, 0) W3 (0, 1) −W1 (1, 0) W2 (1, 1) W3 (0, 1) = 10 + 2 · 3 = 16 Hence E(110) =

8 9

the same value we obtained earlier and referred to as relative information for the two-factor interaction A1 × A2 . If we define the efficiency factor for a BFD as E=

(si − 1)zi E(z) T RL =1− t −1 t −1

(12.35)

z∈Z i

with TRL being the total relative loss defined earlier, we obtain, for both designs of Table 12.9, E = 10 11 . To choose among competing designs, it is therefore in general advisable to consider not only E for each design but also the individual E(z) and then choose the design that is most appropriate for the experiment under consideration.

502


12.5.3 Construction of BFDs We shall now turn to the construction of BFDs or, alternatively, to the construction of EGD-PBIB designs. For some of these methods we shall refer to Chapter 5 and provide here just an illustration, for other methods we shall give a brief description of their special features in the context of factorial experiments. 12.5.3.1 Kronecker Product Designs The method is described in Section 5.5, and we know from Chapter 4 that the Kronecker product of BIBDs is an EGD-PBIB design. As an illustration let us consider Example 5.5 to construct a BFD for a 4 × 3 factorial in blocks of size 6. Using the SDP notation of Section 11.9, the treatment combinations are written in the order (0, 0), (0, 1), (0, 2), (1, 0), (1, 1), (1, 2), (2, 0), (2, 1), (2, 2), (3, 0), (3, 1), (3, 2). The 12 blocks are then as given in Table 12.10. Following our 29 earlier discussion, it is easy to work out that E(1, 0) = 31 36 , E(0, 1) = 36 , and 35 E(1, 1) = 35 36 , and the overall efficiency factor is E = 1 − 396 . 12.5.3.2 GD-PBIB(2) Designs If in an EGD(3)-PBIB design we have λ(10) = λ(11), then the design reduces to a GD-PBIB(2) design with parameters n1 = n(01), n2 = n(10) + n(11) and λ1 = λ(01), λ2 = λ(10) = λ(11). It is therefore possible to obtain a BFD for a t = s1 × s2 factorial if there exists a GD-PBIB design for t = s1 s2 treatments. Table 12.10 4 × 3 Factorial in Blocks of Size 6 Using a Kronecker Product Design Block 1

2

3

4

5

6

00 01 10 11 20 21

00 02 10 12 20 22

01 02 11 12 21 22

00 01 10 11 30 31

00 02 10 12 30 32

01 02 11 12 31 32

7

8

9

10

11

12

00 01 20 21 30 31

00 02 20 22 30 32

01 02 21 22 31 32

10 11 20 21 30 31

10 12 20 22 30 32

11 12 21 22 31 32

Block

503


All we need to do is to arrange the factorial treatment combinations in the rectangular array (0, 0) (1, 0) .. .

(0, 1) (1, 1) .. .

(s1 − 1, 0)

(s1 − 1, 1)

··· ··· .. . ···

(0, s2 − 1) (1, s2 − 1) .. . (s1 − 1, s2 − 1)

and use the association scheme for a GD-PBIB design: Any two treatments in the same row are first associates and any two treatments in different rows are second associates. We can then use any suitable GD-PBIB design to obtain a BFD. This method was proposed by Kramer and Bradley (1957) and Zelen (1958). We shall illustrate this method for the 4 × 3 factorial in blocks of size 6 using design S 27 of Clatworthy (1973). Using the following correspondence between ordinary treatments and factorial treatments, 1

5

9

00

01

02

2

6

10

10

11

12

3

7

11

20

21

22

4

8

12

30

31

32

−→

the BFD is given in Table 12.11 [we note here that this labeling does not agree with that obtained from the SDP of (0 1 2 3) and (0 1 2) but rather it agrees with the labelling for GD-PBIB designs as given by Clatworthy (1973) and that for EGD-PBIB designs]. With λ0 = r = 3, λ1 = 3, λ2 = 1 for the GD-PBIB and hence λ(00) = r = 3, λ(01) = 3, λ(10) = λ(11) = 1 for the EGD-PBIB we can obtain the efficiency factors in the usual way except for one small change. From

Table 12.11 4 × 3 Factorial in Blocks of Size 6 Using a GD-PBIB Block 1

2

3

4

5

6

00 10 01 11 02 12

20 30 21 31 22 32

02 22 00 20 01 21

12 32 10 30 11 31

01 31 02 32 00 30

11 21 12 22 10 20

504


the structure of the GD-PBIB design we know that we can write B0 = I t

! B 1 = Is2 I s2 − I s2 × I s1 = Is2 I s2 × I s1 − I t B 2 = Is2 I s2 × Is1 I s1 − I s1

!

= It I t − Is2 I s2 × I s1 and B 0 = B (00) = D 02 × D 01 B 1 = B (01) = D 12 × D 01 − D 02 × D 01 B 2 = B (10) + B (11) = D 12 × D 11 − D 12 × D 01 Notice that because of the labeling the order of the D matrices is reversed (this is, of course, essential only for D 12 × D 01 ). Hence, in general,

r − λ1 C= r− k

D 02 × D 01 −

λ1 − λ2 1 λ2 D 2 × D 01 − D 12 × D 11 k k

so that, using (12.34) and the properties of the Wi , r − λ1 rE(1, 0) = r − W1 (1, 0)W2 (0, 0) k −

λ1 − λ2 W1 (1, 0)W2 (0, 1) k

r − λ1 rE(0, 1) = r − k and

W1 (0, 0)W2 (1, 0)

λ1 W2 (1, 0)W1 (1, 0) rE(1, 1) = r − k

Then, for the specific example of Table 12.11 we find E(1, 0) =

2 3

E(0, 1) = 1

E(1, 1) = 1

that is, we have full information on the main effect A2 and the interaction A1 × A2 and two-thirds information on the main effect A1 . The GD-PBIB design for t = t1 t2 treatments can, of course, also be used for factorials with more than two factors, that is, for an s1 × s2 × · · · × sn factorial, as long as si1 si2 · · · sim = t1 for any m factors (m < n) and sj1 sj2 · · · sj = t2 for

505


the remaining = n − m factors. Thus the design in Table 12.11 can also be used for the 2 × 2 × 3 factorial in blocks of size 6 with t1 = 4 = s1 s2 = 2 · 2. All we need to do is relabel appropriately the levels of A1 . It follows then immediately that E(100) = E(010) = E(110) =

2 3

and all other E(z) = 1. 12.5.3.3 Other Methods The methods of constructing EGD-PBIB designs given by Aggarwal (1974) and Chang and Hinkelmann (1987) (see Section 5.5) can be used to obtain different systems of confounding for the s1 × s2 × · · · × sn factorial. If the factors are labeled such that s1 ≤ s2 ≤ · · · ≤ sn , then Aggarwal’s method leads to BFDs with block size k = s1 . The method by Chang and Hinkelmann is somewhat more flexible in that it can be used to obtain BFDs with block sizes k = si (i = 1, 2, . . ., n − 1). It leads, however, to disconnected designs, which means that some effects are completely confounded with blocks. Both methods can be modified easily to generate a larger class of BFDs by combining several factors into “superfactors” and thereby changing the order of the levels, that is, some si = s1 becomes the smallest number of levels, or creating new levels, that is, sn may no longer be the largest number of levels. The first modification is useful for the Aggarwal method, the second for the Chang and Hinkelmann method. There exist other ad hoc methods used by Yates (1937b) and Li (1944) for specific asymmetrical factorials. For example, the design for the 22 × 3 factorial given in Table 12.9b can be obtained by combining the levels of factor A3 with treatment combinations for factors A1 and A2 determined by the equations x1 + x2 = 0 : (0, 0), (1, 1) and x1 + x2 = 1 : (1, 0), (0, 1) If we denote {(0, 0), (1, 1)} by α and {(1, 0), (0, 1)} by β, then the design can be written in compact form as follows (Kempthorne, 1952): Block 1

2

3

4

5

6

α0

β0

β0

α0

β0

α0

β1

α1

α1

β1

β1

α1

β2

α2

β2

α2

α2

β2

506


Rather than describe the ad hoc methods in detail, we reproduce some of the more useful designs given by Li (1944) for up to four factors with two, three, and four levels. They are given in Appendix D together with some relevant parameters. Other methods of construction have been proposed by Kishen and Srivastava (1959) and Muller (1966). They consider s × s1 × · · · × sn−1 factorials, where s is a prime or prime power and various block sizes. It is quite obvious that BFDs require in general several replications that may lead to more blocks than is feasible from a practical point of view. If one is willing to give up the property of balancedness, then useful designs can be obtained by a suitable reduction of the number of replications. Such possibilities are mentioned for some situations in Appendix D. Even though these designs are no longer BFDs, they yield full or partial information on all degrees of freedom. This may make these designs more useful than comparable designs obtained by methods discussed in Sections 11.2 and 11.3.

C H A P T E R 13

Fractional Factorial Designs

13.1 INTRODUCTION So far, we have considered possible arrangements of the complete factorial system; that is, we have considered only arrangements in which each individual treatment combination of the factorial set is represented the same number of times. For example, with 4 factors each at 2 levels, we have discussed arrangements of the 16 treatment combinations in blocks of size 2, blocks of size 4, 8 × 8 squares, and so on, but always with the restriction that each of the 16 combinations is to be tested an equal number of times. Such a restriction is practicable with a small number of factors, but, when we consider the case of, say, 10 factors each at 2 levels, it would result in the necessity of testing 1024 combinations or a multiple of these. The main reason for imposing the restriction is that it results in the estimates of effects and interactions having maximum precision and being uncorrelated. Thus, to take a simple example, suppose we are evaluating 2 factors, A and B, each at 2 levels, and test the treatment combination (1), a, b, ab with r1 , r2 , r3 , r4 replications, respectively. Then using the definitions of Section 7.2.1, each effect and interaction is estimated with variance 1 1 1 σe2 1 + + + 4 r1 r2 r3 r4 but the estimates are correlated thus: σe2 1 1 1 1 − − + cov A, B = 4 r1 r2 r3 r4 σe2 1 1 1 1 cov A, AB = − − + + 4 r1 r2 r3 r4

Design and Analysis of Experiments. Volume 2: Advanced Experimental Design By Klaus Hinkelmann and Oscar Kempthorne ISBN 0-471-55177-5 Copyright  2005 John Wiley & Sons, Inc.

507

508

and

FRACTIONAL FACTORIAL DESIGNS

σe2 1 1 1 1 cov B, AB = − + − + 4 r1 r2 r3 r4

These covariances will be zero if r1 = r2 = r3 = r4 . Furthermore, if we fix (r1 + r2 + r3 + r4 ), the total number of observations, we obtain minimum variance for each of the effects and interaction with r1 = r2 = r3 = r4 . As we have seen throughout, the general problem of the design of experiments is that we are given a model and we wish to determine parametric functions of such a model with as low a variance as possible. There are two methods by which this may be accomplished: namely, by choosing the pattern of observations, for example, equal r’s above, or by reducing the error variance σe2 by choice of the error control design. The question we ask in this chapter is whether it is always necessary to test all the factorial combinations equally frequently or whether we can omit some of them. The question is of considerable relevance, for factorial experiments are most appropriate for exploratory research, and in such research the number of possible factors that should be tested is by no means so small as 2 or 3. For example, in research into the possible importance of the various vitamins in the nutrition of an organism, the number of factors that might be used is at least of the order of 10. It is, furthermore, important to use a factorial system because we cannot assume all interactions to be negligible. The testing of 1024 treatment combinations may be virtually impossible from the practical viewpoint. If the whole 1024 combinations were tested, the subdivision of the 1023 treatment comparisons would be Main effects 2-factor interactions 3-factor interactions 4-factor interactions 5-factor interactions 6-factor interactions 7-factor interactions 8-factor interactions 9-factor interactions 10-factor interactions Total

10 45 120 210 252 210 120 45 10 1 1023

It may well be reasonable to assume that high-order interactions are negligible, and, if, for instance, all interactions involving 3 or more factors could be ignored, the testing of the 1024 combinations would give an estimate of error based on about 950 degrees of freedom. The accuracy that would result for the estimation of main effects and interactions (a variance of σe2 /256 where σe2 is the experimental error variance) might well be unnecessarily high. We may then ask what

509

SIMPLE EXAMPLE OF FRACTIONAL REPLICATION

information can be obtained from the testing of a lesser number of combinations. We shall, in fact, find that information on main effects and interactions for all the 10 factors can be obtained under some mild assumptions from the testing of 512, 256, or perhaps even 128 of the total of 1024 combinations. The general process by which this is accomplished is known as fractional replication. The resulting design is called a fractional factorial design, which was introduced first by Finney (1945). For an annotated bibliography of applications papers using some of the concepts developed in the following chapters, see Prvan and Street (2002). 13.2 SIMPLE EXAMPLE OF FRACTIONAL REPLICATION We shall illustrate the concept of a fractional factorial design in terms of a simple, though somewhat unrealistic, example. Suppose we are testing three factors A, B, and C, each of two levels, and for some reason we are able to observe only four of the eight possible treatment combinations. Let these four treatment combinations be a, b, c, and abc. The question then is: What information can we obtain from these four combinations? This question can be answered easily by referring to Table 13.1, which shows the relationship between the true yields of these four treatment combinations and the effects and interactions for the ordinary 23 design. It shows that, for these four treatment combinations, M and 12 ABC have always the same sign, and so have 12 A and 12 BC , 12 B and 12 AC, 12 C, and 12 AB . It will therefore be impossible to separate the mean from the ABC interaction, the A effect from the BC interaction, the B effect from the AC interaction, and the C effect from the AB interaction. In fact, we have the following four estimating relations: 1 4 (a

+ b + c + abc) estimates M + 12 ABC

1 2 (a

− b − c + abc) estimates A + BC

1 2 (−a

+ b − c + abc) estimates B + AB

1 2 (−a

− b + c + abc) estimates C + AB

where a, b, c, abc are now the observed yields. We may say that, with these four observations only, A is completely confounded with BC because there is Table 13.1 Relationship Between True Yields and Effects and Interactions for 23 Design True Yield

M

1 2A

1 2B

1 2 AB

1 2C

1 2 AC

1 2 BC

1 2 ABC

a b c abc

+ + + +

+ − − +

− + − +

− − + +

− − + +

− + − +

+ − − +

+ + + +

510


no possibility of estimating A alone and BC alone but only their sum. We also express this by saying that A is an alias of BC (and vice versa) or that A is aliased with BC . Similarly, B is completely confounded with AC , C with AB, and M with ABC . This confounding, however, need not cause the experimenter any worry if he is prepared to assume that the interactions are negligible: Then the second, third, and fourth comparisons above may be used to estimate the three main effects, and the estimates are uncorrelated. In this simple example, there is not, of course, any possibility of estimating error variance, and we would not, generally, use this design for any practical purpose. It serves, however, to bring out the main idea of fractional replication, namely that a suitably chosen subset of the full factorial set can provide worthwhile information. The above example utilizes a 12 replication of the 23 system because we have tested only four of the eight combinations. We could equally well have used the treatment combinations (1), ab, ac, and bc, and it may be verified that the estimating equations will be M − 12 ABC = 14 [(1) + ab + ac + bc] A − BC = 12 [−(1) + ab + ac − bc] B − AC = 12 [−(1) + ab − ac + bc] C − AB = 12 [−(1) − ab + ac − bc] The confounding among the mean, main effects, and interactions is the same as before, only the functions that can be estimated are different. The dominant feature of this example is the choice of the interaction ABC and the selection of either the treatments entering positively in this interaction or those entering negatively, or alternatively, either those that satisfy the equation x1 + x2 + x3 = 1 mod 2 or those satisfying the equation x1 + x2 + x3 = 0 mod 2, respectively For most purposes we may disregard which half of the treatment combinations we choose because we shall always choose an interaction in such a way that those interactions that are confounded with the effects or interactions we wish to estimate are assumed on the basis of prior knowledge to be negligible. If we use the equality sign to denote “confounded with,” we may write the confounding relations above in the form: A = BC B = AC C = AB

(13.1)

SIMPLE EXAMPLE OF FRACTIONAL REPLICATION

511

and, if by convention we denote M by I , also I = ABC

(13.2)

We note the equalities (13.1) are obtained from (13.2) by regarding the relationship (13.2) as an algebraic identity and I as unity and using multiplication with the rule that the square of any letter is to be replaced by unity. Thus, (13.2) multiplied by A gives A = A2 BC = BC when multiplied by B gives B = AB 2 C = AC and when multiplied by C gives C = ABC 2 = AB In Table 13.2 we show how this design and the alias structure can be obtained using SAS PROC FACTEX. We comment briefly on the input statements and the output: 1. The number of treatment combinations to be used can be specified either by “size design = # of treatment combinations” or “size fraction = denominator of fraction.” 2. The model can either be specified by listing the effects we wish to estimate or by giving the resolution of the design (see Section 13.3.2). 3. The factor confounding rule is equivalent to the defining relation in the sense that C = A ∗ B is formally identical to C 2 = A ∗ B ∗ C and C 2 = I ; operationally C = A ∗ B indicates how from a complete 22 factorial the levels of the factor C are obtained, namely by multiplying the levels of factors A and B into each other when the levels are denoted by −1 and +1. If, instead of a, b, c, abc, we had chosen the treatment combinations that enter positively into AB, namely (1), ab, c, and abc, we can obtain the confounding relationships formally by multiplying each effect or interaction into I = AB

(13.3)

Thus A = A2 B = B C = ABC AC = A BC = BC 2

(13.4)

512


Table 13.2 1/2 Fraction of 23 Factorial With Defining Contrast I = ABC options nodate pageno=1; proc factex; factors A B C; size design=4; model est=(A B C); examine design aliasing confounding; title1 'TABLE 13.2'; title2 '1/2 FRACTION OF 2**3 FACTORIAL'; title3 'WITH DEFINING CONTRAST I=ABC'; run; proc factex; factors A B C; size fraction=2; model res=3; examine design aliasing confounding; run; The FACTEX Procedure Design Points Experiment Number A B C --------------------1 -1 -1 1 2 -1 1 -1 3 1 -1 -1 4 1 1 1 Factor Confounding Rules C = A*B Aliasing Structure A = B*C B = A*C C = A*B

The equations (13.4) we interpret then by saying that A and B are confounded, C and ABC are confounded, and AC and BC are confounded. Of course, (13.3) means that M and AB are confounded. Alternatively, we can say that with the four observations above we can only estimate M + 12 AB, A + B, C + ABC, and

513

FRACTIONAL REPLICATES FOR 2n FACTORIAL DESIGNS

AC + BC. This can be verified easily by using an argument similar to the one we used at the beginning of this section. In each of these two examples we have used a 12 replicate of the full set of factorial combinations. The information obtained from these two designs is quite different: If we can assume that the three factors do not interact, the first replicate is a satisfactory design (apart from the fact that no estimate of error can be made) because we can estimate each of the three main effects A, B, and C. In the second 12 replicate, however, the effect A is confounded with the effect B, and this is generally undesirable. The crucial part of the specification of the design is clearly the choice of the relationship I = ABC or I = AB. This relation is known as the defining relation or identity relationship. Once this relationship is chosen, we can specify the functions of the parameters that can be estimated, and the choice of an identity relationship is based on this fact. Thus, we would typically not use a relationship that results in confounding of main effects with each other. In passing we note that the defining relation can be written in such a way that it identifies uniquely the two complimentary 12 replicates, for example, I = + ABC ,

and

I = − ABC

the first denoting the 12 replicate based on the treatment combinations entering positively into ABC , and the second based on those entering negatively into ABC . The specific form of the defining relation also tells us which functions of the parameters can be estimated; for example, A + BC from the first and A − BC from the second 12 replicate. In either case, A and BC are confounded, and so forth. The totality of all confounding relationships is also referred to as the alias structure of the fractional factorial design. 13.3 FRACTIONAL REPLICATES FOR 2n FACTORIAL DESIGNS It is clear from the discussion in Section 13.2 that for a 12 replicate of a 2n factorial design will generally be determined by the identity relationship of the form I = A1 A2 · · · An

(13.5)

so that main effects and low-order interactions are aliased with interactions of as high an order as possible. The actual choice of the specific 12 replicate based on (13.5) is often made so that it contains the treatment combination (0, 0, . . . , 0), which is often the control. 13.3.1 The

1 2

Fraction

In a practical situation it may be that a 12 replicate is still too large a fraction to be accommodated in a reasonable experimental design. Thus smaller fractions

514


may have to be considered, such as a 21 fraction of a 2n factorial design. We denote such a design by 2n− . The treatment combinations of a 2n− design are obtained by specifying equations: α11 x1 + α12 x2 + · · · + α1n xn = δ1 α21 x1 + α22 x2 + · · · + α2n xn = δ2

(13.6)

.. . α1 x1 + α2 x2 + · · · + αn xn = δ where the α i = (αi1 , αi2 , . . . , αin ) represent independent interactions E αi (i = 1, 2, . . . , ) (see Section 8.2) and the δi (i = 1, 2, . . . , ) are either 0 or 1 mod 2. It follows then that there are exactly 2n− treatment combinations that satisfy (13.6) and these constitute the specified 2n− fractional factorial. We note the following: 1. All treatment combinations satisfying (13.6) enter with the same sign into E αi (i = 1, 2, . . . , ). 2. Any treatment combination that satisfies (13.6) also satisfies any linear combination of the equations in (13.6). 3. Because of conditions 1 and 2 all treatment combinations satisfying (13.6) also enter with the same sign into any GI of the E αi (i = 1, 2, . . . , ). It then follows that the E αi and all their GIs are confounded with (or inseparable from) the mean M, so that the identity relationship is then I = E α1 = E α2 = E α1 +α 2 = E α3 = E α1 +α 3 = E α2 +α3 = E α1 +α 2 +α 3 = · · · = E α1 +α 2 +···+α

(13.7)

that is, (13.7) contains 2 − 1 interactions. The alias structure is obtained by taking an E β not represented in (13.7) and multiplying it into each term in (13.7), that is, E β is confounded with E α1 +β , E α2 +β , E α1 +α 2 +β , E α3 +β , . . . , E α1 +α 2 +...+α +β

(13.8)

There are s = 22− − 1 such interactions, say E β j (j = 1, 2, . . . , s) and each is aliased with s − 1 other interactions according to (13.8). We thus have the complete alias structure for the 2n− design based on (13.7). It is, of course, clear from Eqs. (13.6) that there exist 2 different fractions 2n− depending on the choice of the δi (i = 1, 2, . . . , ). To specify which fraction we are choosing we may attach a plus or minus sign to the independent interactions E αi (i = 1, 2, . . . , ) so that +E αi (−E αi ) means that we take all treatment combinations that enter positively (negatively) into E αi . Regardless of

515


what fraction we choose, as long as it is based on one of the possible sets of Eqs. (13.6), the alias structure remains the same, only the estimable linear functions of effects and interactions change. To illustrate these concepts we consider the following example. 25

Example 13.1 Let n = 5 and = 2, that is, we consider a design or a 25−2 design. We may specify (13.6) as x1 + x2 + x3

1 4

replicate of a

=0 + x4 + x5 = 0

x1

which are satisfied by the treatment combinations (1), bc, abd, acd, abe, ace, de, bcde These treatment combinations also satisfy the equation x2 + x3 + x4 + x5 = 0 and, moreover, these treatment combinations enter negatively into ABC and ADE and hence positively into their GI BCDE. The identity relationship therefore is I = −ABC = −ADE = +BCDE

(13.9)

From (13.9) we then obtain the alias structure given in Table 13.3. It, together with (13.9), also shows which parametric functions are estimable, for example, A − BC − DE + ABCDE , which is defined as 14 [abd + acd + abe + ace − (1) − bc − de − bcde]. Had we selected the eight treatment combinations according to the equations x1 + x2 + x3 x1

=1 + x4 + x5 = 0

x2 + x3 + x4 + x5 = 1 Table 13.3 Alias Structure for 25−2 Design A = BC = DE = BCDE B = AC = ABDE = CDE C = AB = ACDE = BDE D = ABCD = AE = BCE E = ABCE = AD = BCD BD = ACD = ABE = CE BE = ACE = ABD = CD

516


that is, b, c, ad, ae, bde, cde, abce, abcd the identity relationship would have been I = ABC = −ADE = −BCDE

(13.10)

From the alias structure given in Table 13.3 and from (13.10) we see that now, for example, A + BC − DE − ABCDE is estimable. Rather than specifying the identity relationship (13.7) by attaching plus or minus signs to the E αi , we can indicate the specific fractional replicate by writα ing Eδii , where δi is the ith right-hand side in (13.6), so that according to the definition of Eδαii given in Section 7.4 this implies that we select all treatment combinations that satisfy the equation αi1 x1 + αi2 x2 + · · · + αin xn = δi (i = 1, 2, . . . , ). We shall see that this notation is useful when we consider the p n− system (see Section 13.5). 13.3.2 Resolution of Fractional Factorials As has been mentioned earlier the value of fractional replication lies in the fact that often higher-order interactions are negligible. One would therefore consider fractions that result in the confounding of main effects and low-order interactions with higher-order interactions. The extent to which this is possible is expressed clearly by the form of the identity relationship. Basically, the minimum number of factors involved in an interaction included in the identity relationship determines the resolution of the design, that is, the extent to which main effects and loworder interactions are estimable under the assumption that certain higher-order interactions are negligible. Following the terminology introduced by Box and Hunter (1961a, 1961b), we distinguish between the following types of fractional replicates. Resolution III Designs No main effect is confounded with another main effect, but main effects are confounded with 2-factor interactions. Resolution IV Designs No main effect is confounded with another main effect or a 2-factor interaction, but 2-factor interactions are confounded with one another. Resolution V Designs No main effect is confounded with another main effect or 2-factor interaction, and 2-factor interactions are confounded with 3factor interactions.


517

These definitions apply, strictly speaking, only to so-called regular fractional factorial designs for symmetrical factorials. A regular fraction is, for example, a 21 of a 2n factorial as compared to an irregular fraction such as a 23 of a 2n factorial (see, e.g., Addelman, 1961; Webb, 1971). A more general definition was given by Webb (1964). Definition 13.1 A fractional factorial design is of resolution (2R + 1) if it permits the estimation of all effects up to R-factor interactions assuming that all interactions involving more than R factors are negligible (R = 1, 2, . . .). A design is of resolution 2R if it permits the estimation of all effects up to (R − 1)-factor interactions assuming that all interactions involving (R + 1) factors are negligible (R = 2, 3, . . .). As a consequence of this definition, a resolution III design (i.e., R = 1) permits the estimation of main effects assuming that all interactions are negligible, and a resolution IV design (i.e., R = 2) permits the estimation of main effects assuming that all 3-factor and higher-order interactions are negligible. The design given in Example 13.1 is a resolution III design, which we indicate by calling it a 25−2 III design. Note also that the lowest order interaction in the identity relationship involves three factors or, using different terminology, the smallest word (i.e., interaction) contains three letters (i.e., factors). A 2n− IV is given in the following example. Example 13.2 Let n = 6, = 2, and I = ABCD = CDEF = ABEF

(13.11)

The 16 treatment combinations are obtained by solving the equations x1 + x2 + x3 + x4 = 0 x3 + x4 + x5 + x6 = 0 It is clear from (13.11) that main effects are confounded with 3-factor interactions. We can use SAS PROC FACTEX to generate this design. It is given in Table 13.4 The output shows that the defining relationship I = BCDE = ACDF = ABEF was used, which is of the same form as (13.11). As a consequence seven of the fifteen 2-factor interactions are confounded with one other 2-factor interaction and one 2-factor interaction is confounded with two other 2-factor interactions.

518


13.3.3 Word Length Pattern In discussing the resolution of a fractional factorial design, we have referred to the interactions in the defining relationship as words, and the number of factors involved in an interaction is called the word length. Typically, a defining relationship contains several words of possibly different length. This is referred to as the word length pattern (WLP) of the defining relationship. A 2n− fraction may be constructed by using different defining relationships with different WLPs. More precisely, a defining relationship can be characterized with respect to its general properties by giving the distribution of words of different length. Denoting the number of words with in letters, that is, an m-factor interaction, by Lm , we denote the WLP of an identity relationship by WLP(L3 , L4 , . . . , Ln ) We ignore designs with words of length 1 and 2 since they are obviously of no practical interest as the smallest Lm = 0 determines the resolution of the design. For a 2n− design we then have n

Lm = 2 − 1

m=3

For the designs in Example 13.2 and in Table 13.4 we have WLP(0, 3, 0, 0), and it is easy to see that this is the only possible WLP for a 26−2 design. 13.3.4 Criteria for Design Selection In general, however, there may be different WLPs for a 2n− design, that is, R different identity relationships. In such situations the particular WLP can be used to distinguish among competing designs and select the one that is “best.” And best in our context means the largest number of estimable, that is, unconfounded, effects or interactions. For example, for a resolution III we would like the number of main effects that are confounded with 2-factor interactions as small as possible, or expressed differently, the number of main effects not confounded with 2-factor interactions as large as possible. For a resolution IV design we would like to have as few 2-factor interactions confounded with each other. Wu and Chen (1992) called a main effect or 2-factor interaction clear if none of its aliases are main effects or 2-factor interactions, and strongly clear if none of its aliases are main effects, 2-factor interactions or 3-factor interactions (see also Wu and Hamada, 2000). Another criterion to distinguish among competing designs with maximum possible resolution, Rmax say, was introduced by Fries and Hunter (1980).


Table 13.4 1/4 Replicate of 26 Factorial of Resolution IV options nodate pageno=1; proc factex; factors A B C D E F; size fraction=4; model res=4; examine design aliasing confounding; title1 'TABLE 13.4'; title2 '1/4 REPLICATE OF 2**6 FACTORIAL'; title3 'OF RESOLUTION IV'; run; The FACTEX Procedure Design Points Experiment Number A B C D E F -----------------------------------1 -1 -1 -1 -1 -1 -1 2 -1 -1 -1 1 1 1 3 -1 -1 1 -1 1 1 4 -1 -1 1 1 -1 -1 5 -1 1 -1 -1 1 -1 6 -1 1 -1 1 -1 1 7 -1 1 1 -1 -1 1 8 -1 1 1 1 1 -1 9 1 -1 -1 -1 -1 1 10 1 -1 -1 1 1 -1 11 1 -1 1 -1 1 -1 12 1 -1 1 1 -1 1 13 1 1 -1 -1 1 1 14 1 1 -1 1 -1 -1 15 1 1 1 -1 -1 -1 16 1 1 1 1 1 1 Factor Confounding Rules E = B*C*D F = A*C*D Aliasing Structure A B C D E F

519

520


Table 13.4 (Continued ) A*B A*C A*D A*E A*F B*C B*D

= = = = = = =

E*F D*F C*F B*F B*E = C*D D*E C*E

Suppose we have two competing 2n− designs, D1 and D2 say, of maximum resolution and suppose that the WLP are given as D1 : WLP(L3 , L4 , . . . , Ln ) and D2 : WLP(L3 , L4 , . . . , Ln ) Consider then the first i for which Li = Li (Rmax ≤ i ≤ n). If Li < Li , then D1 is said to have less aberration than D2 , otherwise D2 has less aberration. Among all competing designs the one with the smallest Li is said to be the minimum aberration design. As an illustration of these concepts, we consider the following example. Example 13.3 Let n = 8, = 3, and R = IV; that is, we are interested in a 28−3 IV design consisting of 32 runs. Design 1 I = ABCDEF = CDEG = ABFG = BDEH = ACFH = BCGH = ADEFGH with WLP(0, 5, 0, 2, 0, 0,). This design, as constructed by SAS PROC FACTEX, is given in Table 13.5. Inspection of the alias structure shows that all main effects are clear, but not strongly clear. Also, only 4 of the 28 two-factor interactions are clear (actually, strongly clear), all other 2-factor interactions are confounded with other 2-factor interactions. Design 2 I = CDEF = ABDEG = ABCF G = ABCEH = ABDF H = CDGH = EF GH with WLP(0, 3, 4, 0, 0, 0). This design, given in Table 13.6, was constructed by SAS PROC FACTEX using the option “minabs” to produce a minimum


521

Table 13.5 1/8th Fraction of the 28 Factorial of Resolution IV options nodate pageno=1; proc factex; factors A B C D E F G H; size design=32; model res=4; examine design aliasing confounding; title1 'TABLE 13.5'; title2 '1/8TH FRACTION OF THE 2**8 FACTORIAL'; title3 'OF RESOLUTION IV'; run; The FACTEX Procedure Design Points Experiment Number A B C D E F G H --------------------------------------------1 -1 -1 -1 -1 -1 -1 -1 -1 2 -1 -1 -1 -1 1 1 1 1 3 -1 -1 -1 1 -1 1 1 1 4 -1 -1 -1 1 1 -1 -1 -1 5 -1 -1 1 -1 -1 1 1 -1 6 -1 -1 1 -1 1 -1 -1 1 7 -1 -1 1 1 -1 -1 -1 1 8 -1 -1 1 1 1 1 1 -1 9 -1 1 -1 -1 -1 1 -1 1 10 -1 1 -1 -1 1 -1 1 -1 11 -1 1 -1 1 -1 -1 1 -1 12 -1 1 -1 1 1 1 -1 1 13 -1 1 1 -1 -1 -1 1 1 14 -1 1 1 -1 1 1 -1 -1 15 -1 1 1 1 -1 1 -1 -1 16 -1 1 1 1 1 -1 1 1 17 1 -1 -1 -1 -1 1 -1 -1 18 1 -1 -1 -1 1 -1 1 1 19 1 -1 -1 1 -1 -1 1 1 20 1 -1 -1 1 1 1 -1 -1 21 1 -1 1 -1 -1 -1 1 -1 22 1 -1 1 -1 1 1 -1 1 23 1 -1 1 1 -1 1 -1 1 24 1 -1 1 1 1 -1 1 -1 25 1 1 -1 -1 -1 -1 -1 1 26 1 1 -1 -1 1 1 1 -1 27 1 1 -1 1 -1 1 1 -1 28 1 1 -1 1 1 -1 -1 1 29 1 1 1 -1 -1 1 1 1 30 1 1 1 -1 1 -1 -1 -1 31 1 1 1 1 -1 -1 -1 -1 32 1 1 1 1 1 1 1 1

522


Table 13.5 (Continued ) Factor Confounding Rules F = A*B*C*D*E G = C*D*E H = B*D*E Aliasing Structure A B C D E F G H A*B A*C A*D A*E A*F A*G A*H B*C B*D B*E B*H C*D C*E D*F E*F

= F*G = F*H

= = = = = = = = =

B*G = C*H B*F C*F G*H E*H D*H C*G = D*E E*G D*G

aberration design (L4 = 3 in D2 vs. L4 = 5 in D1 ). All main effects are clear, of course, but only A and B are strongly clear. Now thirteen 2-factor interactions are clear, but none of them are strongly clear. Intuitively, minimum aberration designs should be the best designs with respect to the estimation of 2-factor interactions. However, Wu and Hamada (2000) in their listing of useful 2n− designs give a few examples of resolution IV designs, where a judicious choice of the identity relationship leads to a design with more clear 2-factor interactions than the corresponding minimum aberration design (see also Wu and Wu, 2002). Hence, both concepts should be used when selecting a design for a particular application. Still other criteria have been introduced to further enable the search for and study of the properties of “desirable” designs, such as estimation capacity (Cheng, Steinberg, and Sun, 1999) and estimation index (Chen and Cheng, 2004). In


523

Table 13.6 1/8th Fraction of the 28 Factorial of Resolution IV and Minimum Aberration options nodate pageno=1; proc factex; factors A B C D E F G H; size design=32; model res=4/minabs; examine design aliasing confounding; title1 'TABLE 13.6'; title2 '1/8TH FRACTION OF THE 2**8 FACTORIAL'; title3 'OF RESOLUTION IV AND MINIMUM ABERRATION'; run; The FACTEX Procedure Design Points Experiment Number A B C D E F G H --------------------------------------------1 -1 -1 -1 -1 -1 -1 1 1 2 -1 -1 -1 -1 1 1 -1 -1 3 -1 -1 -1 1 -1 1 -1 1 4 -1 -1 -1 1 1 -1 1 -1 5 -1 -1 1 -1 -1 1 1 -1 6 -1 -1 1 -1 1 -1 -1 1 7 -1 -1 1 1 -1 -1 -1 -1 8 -1 -1 1 1 1 1 1 1 9 -1 1 -1 -1 -1 -1 -1 -1 10 -1 1 -1 -1 1 1 1 1 11 -1 1 -1 1 -1 1 1 -1 12 -1 1 -1 1 1 -1 -1 1 13 -1 1 1 -1 -1 1 -1 1 14 -1 1 1 -1 1 -1 1 -1 15 -1 1 1 1 -1 -1 1 1 16 -1 1 1 1 1 1 -1 -1 17 1 -1 -1 -1 -1 -1 -1 -1 18 1 -1 -1 -1 1 1 1 1 19 1 -1 -1 1 -1 1 1 -1 20 1 -1 -1 1 1 -1 -1 1 21 1 -1 1 -1 -1 1 -1 1 22 1 -1 1 -1 1 -1 1 -1 23 1 -1 1 1 -1 -1 1 1 24 1 -1 1 1 1 1 -1 -1 25 1 1 -1 -1 -1 -1 1 1 26 1 1 -1 -1 1 1 -1 -1 27 1 1 -1 1 -1 1 -1 1 28 1 1 -1 1 1 -1 1 -1 29 1 1 1 -1 -1 1 1 -1 30 1 1 1 -1 1 -1 -1 1 31 1 1 1 1 -1 -1 -1 -1 32 1 1 1 1 1 1 1 1

524


Table 13.6 (Continued ) Factor Confounding Rules F = C*D*E G = A*B*D*E H = A*B*C*E Aliasing Structure A B C D E F G H A*B A*C A*D A*E A*F A*G A*H B*C B*D B*E B*F B*G B*H C*D C*E C*F C*G C*H E*G E*H

= = = = = = =

E*F = G*H D*F D*E D*H D*G F*H F*G

many cases application of these criteria leads to the same design. For a detailed discussion of the relationship among these criteria, we refer the reader to Chen and Cheng (2004). 13.4 FRACTIONAL REPLICATES FOR 3n FACTORIAL DESIGNS The basic ideas for the 2n− design can be extended easily to the 3n− design, that is, the 31 fraction of the 3n factorial. To obtain the treatment combinations

525


for such a fraction we have to specify an identity relationship containing independent interactions. If we denote these interactions by E α1 , E α2 , . . . , E α , we can then write I = Eδα11 = Eδα22 = · · · = Eδα

(13.12)

The treatment combinations are obtained by solving the equations (mod 3) α11 x1 + α12 x2 + · · · + α1n xn = δ1 α21 x1 + α22 x2 + · · · + α2n xn = δ2 .. .

(13.13)

α1 x1 + α2 x2 + · · · + αn xn = δ For any choice of δ1 , δ2 , . . . , δ there will be exactly 3n− treatment combinations satisfying (13.13). Each such set, and there will be 3 possible sets, constitutes a 3n− fractional factorial. If, for some reason, we want the fraction to contain the “control,” that is, the treatment combination (0, 0, . . . , 0), then we choose δ1 = δ2 = · · · = δ = 0. We know, of course, that any treatment combination x = (x1 , x2 , . . . , xn ) satisfying Eqs. (13.13), that is, α i x = δi (i = 1, 2, . . . , ), also satisfies the equations

λi α i

i=1

x=

λi δi

(13.14)

i=1

where λi = 0, 1, 2(i = 1, 2, . . . , ). For of λ1 , λ2 , . . . , λ (not all a given set zero) let us denote λi α i by α ∗ and λi δi by δ ∗ . Thus (13.14) becomes (α ∗ ) x = δ ∗

(13.15)

∗

Each α ∗ corresponds to an effect E α and the treatment combinations satisfying ∗ (13.15) belong to the component Eδα∗ with δ ∗ = 0, 1, 2. It is not difficult to see that, using our convention that the first nonzero element in α ∗ is equal to one, there are altogether (3 − 1)/2 such α ∗ including α 1 , α 2 , . . . , α . And it is obvious that the α ∗ other than α 1 , α 2 , . . . , α correspond to all possible GIs among E α1 , E α2 , . . . , E α . Hence the complete identity relationship is obtained by adding to (13.12) all the GIs, that is, +α 2 I = Eδα11 = Eδα22 = · · · = Eδα = Eδα11+δ 2 +2α 2 +...+2α +2α 2 = Eδα11+2δ = · · · = Eδα11+2δ 2 2 +...+2δ

(13.16)

526


In order to obtain the treatment combinations for the 3n− design, we need to know only (13.12) and solve Eqs. (13.13), but in order to obtain the alias structure we need to use the entire set of interactions in the identity relationship as given in (13.16). Suppose E β is an effect not contained in (13.16). We know that because of orthogonality (see Chapter 10) exactly 13 of the treatment combinations satisfying (13.13) also satisfy the equation β1 x1 + β2 x2 + · · · + βn xn = δ

(13.17)

for δ = 0, 1, 2 and each treatment combination satisfies only one of the three equations. Then E β is defined, as with the full set of 3n treatment combinations, by contrasts of the form β

β

β

c0 E0 + c1 E1 + c2 E2

(13.18)

β

with c0 + c1 + c2 = 0, where Eδ is obtained (apart from the overall mean) from the 3n−−1 treatment combinations satisfying (13.17) for δ = 0, 1, 2. Let us take now any of the interactions in (13.16), say Eδαii , and consider the GI between E β and E αi , denoted by E β+αi . A contrast belonging to this effect is given by β+α i

c0 E 0

β+α i

where the leading mean in Eδ satisfying

β+α i

+ c1 E 1

β+α i

+ c2 E 2

is obtained from the treatment combinations

(β1 + αi1 ) x1 + (β2 + αi2 ) x2 + · · · + (βn + αin ) xn = δ

(13.19)

We know that αi1 x1 + αi2 x2 + · · · + αin xn = δi

(13.20)

Substituting (13.20) in (13.19) yields β1 x1 + β2 x2 + · · · + βn xn = δ − δi Now δ − δi takes on the values 0, 1, 2 (mod 3), and it is therefore obvious that a contrast belonging to E β is also a contrast belonging to E β+αi , that is, E β and E αi are confounded. The same argument holds for the other GI between E β and E αi , namely E β+2αi . This then explains how we use (13.16) to obtain the alias structure for the 3n− design: Any effect E β not in (13.16) is confounded with all GIs obtained by formally multiplying E β into E γ and E 2γ where E γ is any effect listed in (13.16), that is, E β is confounded with E β+γ and E β+2γ . We illustrate this below.

527


Example 13.4 Consider the 34−2 design with the identity relationship given by I = ABC 20 = ACD 0 = AB 2 D02 = BCD 20

(13.21)

where ABC 2 and ACD are the two independent interactions of (13.12) and AB 2 D 2 , BCD 2 are their GIs. The treatment combinations are obtained from the equations x1 + x2 + 2x3 =0 x1

+ x3 + x4 = 0

They are 0000 1011 2022

0112 1120 2101

0221 1202 2210

The alias structure is given by multiplying each effect not contained in (13.21) into each interaction in (13.21) and its square (ignoring the subscripts). We obtain (using reduction mod 3): A = A(ABC 2 ) = A(ABC 2 )2 = A(ACD) = A(ACD)2 = A(AB 2 D 2 ) = A(AB 2 D 2 )2 = A(BCD 2 ) = A(BCD 2 )2 = AB 2 C = BC 2 = AC 2 D 2 = CD = ABD = BD = ABCD 2 = AB 2 C 2 D Similarly B = AB 2 C 2 = AC 2 = ABCD = AB 2 CD = AD 2 = ABD 2 = BC 2 D = CD 2 C = AB = ABC = AC 2 D = AD = AB 2 CD 2 = AB 2 C 2 D 2 = BC 2 D 2 = BD 2 D = ABC 2 D = ABC 2 D 2 = ACD 2 = AC = AB 2 = AB 2 D = BC = BCD Together with the effects in (13.21) these account for all the effects of a 34 factorial. We recognize, of course, that this design is a resolution III design, which allows the estimation of all main effects only if all interactions are negligible. An equivalent design, using SAS PROC FACTEX, is given in Table 13.7. The factor confounding rule states that the levels of C are obtained as twice the levels of A plus twice the levels of B, and that the levels of D are obtained as the levels of A plus twice the levels of B and reduction mod 3. The output shows

528


Table 13.7 1/9 Fraction of the 34 Factorial of Resolution III options nodate pageno=1; proc factex; factors A B C D/nlev=3; size fraction=9; model res=3; examine design aliasing confounding; title1 'TABLE 13.7'; title2 '1/9 FRACTION OF THE 3**4 FACTORIAL'; title3 'OF RESOLUTION III'; run; The FACTEX Procedure Design Points Experiment Number A B C D --------------------------1 -1 -1 -1 -1 2 -1 0 1 1 3 -1 1 0 0 4 0 -1 1 0 5 0 0 0 -1 6 0 1 -1 1 7 1 -1 0 1 8 1 0 -1 0 9 1 1 1 -1 Factor Confounding Rules C = (2*A) + (2*B) D = A + (2*B) Aliasing Structure A (2*A) B (2*B) C (2*C) D (2*D)

= = = = = = = =

(2*B) + (2*C) = B + D = C + (2*D) B + C = (2*B) + (2*D) = (2*C) + D (2*A) + (2*C) = A + (2*D) = C + D A + C = (2*A) + D = (2*C) + (2*D) (2*A) + (2*B) = A + D = B + (2*D) A + B = (2*A) + (2*D) = (2*B) + D A + (2*B) = (2*A) + C = B + (2*C) (2*A) + B = A + (2*C) = (2*B) + C

that the levels of C are indeed computed in this fashion, but the levels of D are obtained by subtracting 1 from the results of the rule given above (presumably this is done to include the control in the fraction). The alias structure shows that, just as in Example 13.4, the main effect A is confounded with 2 d.f. of each of the 2-factor interactions B × C, B × D, and C × D (and higher-order interactions).

GENERAL CASE OF FRACTIONAL REPLICATION

529

The notions of clear effects and minimum aberration design can be carried over in an obvious fashion to 3n− designs. A listing of useful designs is given by Wu and Hamada (2000). 13.5 GENERAL CASE OF FRACTIONAL REPLICATION 13.5.1 Symmetrical Factorials From our discussion above it is quite obvious how the concept of fractional replication can be extended to the general case, that is, to the 1/s fraction of the s n factorial where s is a prime or prime power. All we need to do is specify independent interactions, say E α1 , E α2 , . . . , E α and the s n− treatment combinations are then obtained by solving the equations αi1 x1 + αi2 x2 + · · · + αin xn = δi (i = 1, 2, . . . , ) where αij , xj , δi are elements in GF(s) (i = 1, 2, . . . , ; j = 1, 2, . . . , n). The identity relationship is given by I = E α1 = E α2 = · · · = E α = E α1 +α 2 = E α1 +u2 α 2 = · · · = E α1 +us−1 α2 = · · · = E α1 +us−1 α2 +···+us−1 α

(13.22)

[uj ∈ GF(s)] that is, (13.22) contains the independent interactions and all their GIs. It follows then that an effect E α is aliased with all the GIs between E α and the interactions in (13.22). 13.5.2 Asymmetrical Factorials The basic ideas of fractional replication for symmetrical factorial experiments as presented above can be extended and applied also to asymmetrical factorials. n Suppose we want to choose a fraction of an s1n1 × s2n2 × · · · × s1 q factorial. An obvious and easy method would be to consider a 1 − in − s11 s22 , . . . , sq q replin − cate consisting of s1n1 −1 , s2n2 −2 , . . . , sq q q treatment combinations. The idea, of course, is to consider each of the q symmetrical component systems separately, obtain a 1/sii fraction of the sini factorial (i = 1, 2, . . . , q ) and then combine those fractions as a Kronecker product design. More specifically, let x i = (x i1 , x i2 , . . . , x iνi ) denote the νi = sin−i treatment combinations for the fractional factorial design of the sini factorial, abbreviated FFD(si , ni , i ), for i = 1, 2, . . . , q, with x ij = (xij 1 , xij 2 , . . . , xij ni ) and xij k the level of the kth factor in the j th treatment combination of FFD(si , ni , i ). Then the treatment combinations for the fractional replicate of the asymmetrical factorial s1n1 × s2n2 × n · · · × sq q are given by the symbolic direct product (see

qSection 11.4.1) x 1 ⊗ x 2 ⊗ · · · ⊗ x q . Symbolically, we may write this also as i=1 × FFD(si , ni , i ).

530


Now each FFD (si , ni , i ) is based on an identity relationship, Ii say, containing i independent interactions and all their GI’s. Thus Ii determines the alias structure for FFD(si , ni , i ) involving the ni factors in the sini factorial. We simply apply the rules given above. In addition, we have confounding among mixed interactions, that is, interactions involving factors from different symmetrical factorials. We note here that the alias structure for mixed interactions may be more advantageous than those for interactions in the symmetrical component systems, advantageous in the sense of being confounded with higher-order interactions that may be assumed negligible. We shall illustrate these ideas with the following example. Example 13.5 Consider a 16 replicate of the 23 × 33 factorial; that is, q = 2, s1 = 2, n1 = 3, 1 = 1, s2 = 3, n2 = 3, 2 = 1. Let A1 , A2 , A3 and B1 , B2 , B3 denote the factors for the two symmetrical component systems, respectively. To obtain FFD(2, 3, 1) we use I1 = (A1 A2 A3 )0 and for FFD(3, 3, 1) we use I2 = (B1 B2 B32 )0 It follows that



0 1 x1 =  1 0 and



0 1  2  0  x2 =  1  2 0  1 2

0 1 0 1

 0 0  1 1

0 2 1 1 0 2 2 1 0

 0 0  0  1  1  1 2  2 2

The 36 treatment combinations are obtained by combining every treatment combination in x 1 with every treatment combination in x 2 . The alias structure for FFD(2, 3, 1) is given by A1 = A2 A3 A2 = A1 A3 A3 = A1 A2


531

and for FFD(3, 3, 1) by B1 = B1 B22 B3 = B2 B32 B2 = B1 B22 B32 = B1 B32 B3 = B1 B2 = B1 B2 B3 B1 B22 = B1 B3 = B2 B3 The alias structure for mixed interactions is illustrated by A1 × B1 = A1 × B1 B22 B3 = A1 × B2 B32 = A2 A3 × B1 = A2 A3 × B1 B22 B3 = A2 A3 × B2 B32 This shows that the 2-factor interaction A1 × B1 is confounded with 3-factor or higher-order interactions. An isomorphic design using SAS PROC FACTEX is given in Table 13.8. 13.5.3 General Considerations From a practical point of view highly fractionated designs may be the only feasible designs, requiring only a relatively small number of treatment combinations. This has, of course, fairly severe consequences as each effect will be confounded with a “large” number of other effects. For such a design to be useful the identity relationship must be chosen carefully in order to avoid having effects of interest confounded with each other. For example, if main effects and some or all 2-factor interactions are considered to be important, then certainly main effects should not be confounded with other main effects and 2-factor interactions. If only some 2-factor interactions are important, then confounding of certain 2-factor interactions is permissible, but if all 2-factor interactions are important, then they must only be confounded with 3-factor or higher-order interactions. In other words, resolution IV or V designs are called for. It is obvious that the more effects we want to estimate the more treatment combinations we need. It is typical that for experiments having to use fractional factorial designs the two requirements of a limited number of treatment combinations and of information about a specified number of effects are often in conflict with each other and compromises will have to be made. It is therefore important to be aware of methods of constructing fractional factorial designs of minimal size to achieve certain objectives such as those mentioned above. This leads also to consideration of fractional replicates other than the regular fractions discussed so far. A general discussion of the structures and properties of fractional factorial designs is given by Raktoe, Hedayat, and Federer (1981), and a description of construction methods for certain fractional factorial designs is provided by Dey (1985). We shall discuss some of these issues in Section 13.6.

532


Table 13.8 1/6 Fraction of the 23 × 33 Factorial Using a Kronecker Product Design options nodate pageno=1; proc factex; factors A1 A2 A3; size design=4; model res=3; examine design confounding aliasing; output out=two; title1 'TABLE 13.8'; title2 '1/6 FRACTION OF THE 2**3×3**3 FACTORIAL'; title3 'USING A KRONECKER PRODUCT DESIGN'; run; factors B1 B2 B3/nlev=3; size design=9; model res=3; examine design confounding aliasing; output out=three designrep=two; run; proc print data=three; run; The FACTEX Procedure Design Points Experiment Number A1 A2 A3 -------------------1 -1 -1 1 2 -1 1 -1 3 1 -1 -1 4 1 1 1 Factor Confounding Rules A3 = A1*A2 Aliasing Structure A1 = A2*A3 A2 = A1*A3 A3 = A1*A2

533


Table 13.8 (Continued ) Design Points Experiment Number B1 B2 B3 --------------------1 -1 -1 -1 2 -1 0 1 3 -1 1 0 4 0 -1 1 5 0 0 0 6 0 1 -1 7 1 -1 0 8 1 0 -1 9 1 1 1 Factor Confounding Rules B3 = (2*B1) + (2*B2) Aliasing Structure B1 = (2*B2) + (2*B3) (2*B1) = B2 + B3 B2 = (2*B1) + (2*B3) (2*B2) = B1 + B3 B3 = (2*B1) + (2*B2) (2*B3) = B1 + B2 (2*B1)+ B2 = B1 + (2*B3) = (2*B2) + B3 B1 + (2*B2) = (2*B1) + B3 = B2 + (2*B3) Obs

A1

A2

A3

B1

B2

B3

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19

-1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 1

-1 -1 -1 -1 -1 -1 -1 -1 -1 1 1 1 1 1 1 1 1 1 -1

1 1 1 1 1 1 1 1 1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1

-1 -1 -1 0 0 0 1 1 1 -1 -1 -1 0 0 0 1 1 1 -1

-1 0 1 -1 0 1 -1 0 1 -1 0 1 -1 0 1 -1 0 1 -1

-1 1 0 1 0 -1 0 -1 1 -1 1 0 1 0 -1 0 -1 1 -1

534


Table 13.8 (Continued ) 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

-1 -1 -1 -1 -1 -1 -1 -1 1 1 1 1 1 1 1 1 1

-1 -1 -1 -1 -1 -1 -1 -1 1 1 1 1 1 1 1 1 1

-1 -1 0 0 0 1 1 1 -1 -1 -1 0 0 0 1 1 1

0 1 -1 0 1 -1 0 1 -1 0 1 -1 0 1 -1 0 1

1 0 1 0 -1 0 -1 1 -1 1 0 1 0 -1 0 -1 1

13.5.4 Maximum Resolution Design The discussion in Section 13.6 will allow us to make statements about the minimum size of a design for a specified resolution. An equally important question arises if we are given a fixed number, say N , of runs and we want to know what is the maximum resolution that can be achieved under this constraint. It is clear, of course, that a 1/s fraction of the s n factorial has resolution R = n. For the general case of the s n− fractional factorial an upper bound of the resolution was given by Fujii (1976). We state his result in the following theorem without proof. Theorem 13.1 (Fujii, 1976) Let M = (s − 1)/(s − 1) and n = qM + m, with q ≥ 0 integer. Then, for ≥ 2, the maximum resolution R (n, s) of the s n− design (with s prime or prime power) is given by R (n, s) = s −1 q

if m = 0, 1 ≤ s −1 q + s −2 (s − 1)(m − 1)/(s −1 − 1) if m = 2, 3, . . . , s −1 − 1 ≤ s −1 q + [(s − 1)m/s] if m = s −1 , s −1 + 1, . . . , M − 1

where [x] is the greatest integer not exceeding x. In Table 13.9 we present Rmax values for s = 2, 3, 4, 5 and 4 ≤ n ≤ 10, 2 ≤ ≤ 7 with the number of runs ≤ 256. An asterisk indicates that Rmax < R (n, s)

535


Table 13.9

Maximum Resolution s n− Designs

s

n

Rmax

2

5 6 6 7 7 7 8 8 8 9 9 9 10 10 10 10 4 5 6 6 7 7 7 8 8 8 9 9 9 10 10 10 4 5 5 6 6 7 7 8 8 9 9 10 10

2 2 3 2 3 4 2 3 4 3 4 5 3 4 5 6 2 2 2 3 2 3 4 3 4 5 4 5 6 5 6 7 2 2 3 2 3 3 4 4 5 5 6 6 7

3 4 3 4 4 3 5 4 4 4 4 3∗ 5 4∗ 4 3∗ 3 3 4 3 5 4 3∗ 5 4 3∗ 5 3∗ 3∗ 5∗ 3∗ 3∗ 3 4 3 4 4 4 3∗ 4∗ 3∗ 4∗ 3∗ 4∗ 3∗

3

4

Number of Runs 8 16 8 32 16 8 64 32 16 64 32 16 128 64 32 16 9 27 81 27 243 81 27 243 81 27 243 81 27 243 81 27 16 64 16 256 64 256 64 256 64 256 64 256 64

536


Table 13.9

(Continued )

s

n

Rmax

5

4 5 5 6 6 7 8 9 10

2 2 3 3 4 4 5 6 7

3 4 3 4 3∗ 3∗ 3∗ 3∗ 3∗

Number of Runs 25 125 25 125 25 125 125 125 125

as given above [it shows that R (n, s) is not particularly good for larger values of , a fact also observed by Fries and Hunter (1980) who provided a different bound for s = 2]. 13.6 CHARACTERIZATION OF FRACTIONAL FACTORIAL DESIGNS OF RESOLUTION III, IV, AND V 13.6.1 General Formulation Recall from (12.15) that each true treatment response, a(v), for a s1n1 × s2n2 × n · · · × sq q factorial with s1 < s2 < · · · < sq can be expressed in terms of the mean response, a component from each pure partition and a component from each mixed partition. Of interest here are only all components from pure partitions belonging to main effects and 2-factor interactions and all components from mixed partitions belonging to 2-factor interactions. As an illustration we shall consider the 22 × 32 factorial. Let A1 , A2 denote the two factors with two levels each, and B1 , B2 the two factors with three levels. Recall from Example 12.4 that all arithmetic has to be done in R(6) and that the levels of A1 , A2 are denoted by 0, 3 and those of B1 , B2 by 0, 4, 2. To simplify the notation we shall write the treatment combination v = (v11 , v12 , v21 , v22 ) as v = (v 1 , v 2 ) where v 1 = (v11 , v12 ) , v 2 = (v21 , v22 ) . We then have a(v) = a(v 1 , v 2 )= + (A31 )(30)v 1 + (A32 )(03)v 1 + (A31 A32 )(33)v1 + (B14 )(40)v2 + (B24 )(04)v 2 + (B14 B24 )(44)v2 + (B14 B22 )(42)v 2 + (A31 B14 )∗(3040)v + (A31 B24 )∗(3004)v + (A32 B14 )∗(0340)v + (A32 B24 )∗(0304)v

(13.23)

CHARACTERIZATION OF FRACTIONAL FACTORIAL DESIGNS OF RESOLUTION III, IV, AND V

537

Alternatively, and equivalently, we can express a(v) in terms of the traditional model for factorial structures as a(v) = µ + αi + βj + (αβ)ij + γk + δ +(γ δ)k + (αγ )ik + (βγ )j k +(αδ)i + (βδ)j

(13.24)

with i, j = 0, 1, k, = 0, 1, 2. Models of the type (13.24) are used in SAS, for example. We ignore 3-factor and higher-order interactions since those are assumed to be negligible for resolution III, IV, and V designs. In addition, for resolution III designs 2-factor interactions are also negligible, and we are interested only in estimating the mean and main effects. For resolution IV designs we are mainly interested in estimating main effects, but the mean and 2-factor interactions are assumed to be unknown (and not necessarily zero). For resolution V designs we wish, of course, to estimate the mean, main effects, and 2-factor interactions. All three situations can be expressed in very general terms in the form of a three-part linear model, the first part referring to the mean, the second part referring to the main effects, and the third part to the 2-factor interactions. We formulate this as follows (see Raktoe, Hedayat, and Federer, 1981). Let D denote an FFD consisting of N treatment combinations (not necessarily all distinct) and let a D denote the N × 1 vector of the true treatment responses associated with D. Further, let M denote the mean, E 1 the q1 × 1 vector of main effect components and E 2 the q2 × 1 vector of 2-factor interaction components [using either (13.23) or (13.24)], with E 1 and E 2 conformable to D and the factorial from which D was obtained. We then write, analogously to (12.15), 

aD

 M = (I, X1D , X 2D ) E 1  E2

(13.25)

where X1D and X 2D are N × q1 and N × q2 incidence matrices, respectively. If y D represents the N × 1 vector of observations corresponding to a D , then we write (13.25) alternatively and in a more familiar form as 

 M E[y D ] = (I, X 1D , X2D ) E 1  E2

(13.26)

In order to determine whether a given FFD is of resolution III, IV, or V it is sufficient to consider, instead of D, the unreplicated design D0 , which contains

538


the same treatment combinations as D except that each treatment combination is represented only once. Suppose D0 contains N0 treatment combinations. We then replace (13.26), using obvious notation, by   M E[y D0 ] = (I, X 1D0 , X2D0 ) E 1  (13.27) E2 We shall now consider (13.27) and the resulting NE for each of the three situations. 13.6.2 Resolution III Designs For a resolution III design we have, of course, E 2 = 0 and (13.27) reduces to the two-part model (dropping the subscript D0 for convenience) M E(y) = (I, X 1 ) (13.28) E1 and the NE are of the form M Iy I I I X 1 = X1 y X1 I X1 X1 E1

(13.29)

(i = Suppose we have n factors A1 , A2 , . . . , An and factor Ai has si levels 1, 2, . . . , n), where the si need not be all different. Then E 1 has q1 = i si components, but for factor Ai the si components add to zero (i = 1, 2, . . . , n). Hence E 1 contains only i (si − 1) = q1 − n independent components. It follows then also that the coefficient matrix in (13.29) is not of full rank. A solution to (13.29) can be written as

− M I I I X 1 Iy = (13.30) I X X y X X E1 1 1 1 1 or, if we write X ∗1 = (I, X1 ),

M 1 E

∗ − ∗ = X∗ X1 y 1 X1

(13.31)

It is obvious then that for the mean and all main effects to be estimable we must have ∗ rank X ∗ (13.32) 1 X 1 = 1 + q1 − n Condition (13.32) can in certain situations be achieved with N0 = 1 + q1 − n treatment combinations, and we shall give some examples later (see Chapter 14).


539

For such a design no d.f. are available for estimating the variance σe2 in the context of the ANOVA. The design D0 is then referred to as a u¨ saturated design. If we use a design D with N runs where N > 1 + q1 − n and (13.32) is satisfied, then σe2 can, of course, be estimated with N − (1 + q1 − n) d.f. Using ∗ − (X ∗ 1D X 1D ) as the variance–covariance matrix we can then, within the factor space of the experiment, compare different treatments, for example, a(v) − a(w) for some treatment combinations v and w using the main effects model (13.28). For the analysis of a saturated design D0 see Section 13.9. 13.6.3 Resolution IV Designs For this case it is convenient to rewrite (13.27) (dropping again the subscript D0 ) as

E1 (13.33) E y) = (X 1 , X∗2 E ∗2 with X∗2

= (I, X2 )

E ∗2

M = E2

The reason to include M in E ∗2 , the vector of nuisance parameters, is mostly one of mathematical convenience. It simply means that we are not interested in estimating the mean just as we are not interested in estimating 2-factor interactions. A plausible argument for proceeding in this fashion was given by Raktoe, Hedayat, and Federer (1981) who point out that because of the possible presence of 2-factor interactions we cannot use (13.27) in the same way as we used (13.28) to make comparisons among treatments. Our only aim here can be to estimate main effects and draw conclusions from that concerning the importance of the various factors. The important difference between estimating main effects from a resolution IV design as compared to a resolution III design is, of course, that in a resolution IV design 3-factor interactions are assumed to be negligible, that is, main effects are clear, whereas in a resolution III design 2-factor interactions are also assumed to be negligible. Returning now to model (13.33) the NE are

X1 X1 X1 X∗2 E1 X1 y (13.34) = ∗ ∗ ∗ ∗ X∗ X2 X1 X2 X2 E2 2y Using results of Chapter I.4 the RNE for E 1 are given by ∗ − ∗ E1 X X X X1 X 1 − X1 X∗2 X∗ 1 2 2 2 ∗ − ∗ = X 1 − X1 X∗2 X∗ X X 2 2 2 y

(13.35)

540


or CE 1 = Q

(13.36)

where C and Q are defined in an obvious way. For all main effects to be estimable we must have ∗ − ∗ X X X (si − 1) = q1 − n (13.37) rank X1 X 1 − X1 X∗2 X∗ 2 2 2 1 = i

To illustrate these ideas we consider the following example. Example 13.6 Suppose we have four factors, A, B, C, F , each at two levels and we have a 12 replicate D0 = {(0 0 0 0), (0 1 1 0), (1 0 1 0), (1 1 0 0), (0 0 1 1), (0 1 0 1), (1 0 0 1), (1 1 1 1)} with eight treatment combinations. Notice that D0 is based on the identity relationship I = ABCF . So we already know that D0 is a resolution IV design. We now have  A0 1   1   0  0 X1 =    1  1   0 0 X∗2 =  M       

1 1 1 1 1 1 1 1

AB0 1 0 0 1 1 0 0 1

AB1 0 1 1 0 0 1 1 0

A1 0 0 1 1 0 0 1 1

AC0 1 0 1 0 0 1 0 1

      X1 X1 =     

B0 1 0 1 0 1 0 1 0

AC1 0 1 0 1 1 0 1 0

4 0 2 2 2 2 2 2

B1 0 1 0 1 0 1 0 1

AF0 1 1 0 0 0 0 1 1

0 4 2 2 2 2 2 2

2 2 4 0 2 2 2 2

C0 1 0 0 1 0 1 1 0

AF1 0 0 1 1 1 1 0 0

2 2 0 4 2 2 2 2

C1 0 1 1 0 1 0 0 1

BC0 1 1 0 0 0 0 1 1

2 2 2 2 4 0 2 2

F0 1 1 1 1 0 0 0 0

BC1 0 0 1 1 1 1 0 0

2 2 2 2 0 4 2 2

F1 0 0 0 0 1 1 1 1

           

BF0 1 0 1 0 0 1 0 1

2 2 2 2 2 2 4 0

= 4I 4 × I 2 + 2(I4 I4 − I 4 ) × I2 I2

2 2 2 2 2 2 0 4

BF1 0 1 0 1 1 0 1 0

          

CF0 1 0 0 1 1 0 0 1

CF1 0 1 1 0 0 1 1 0

       


∗ X∗ 2 X2 =

4I12 

∗ −  (X ∗ 2 X2 ) =

4 I12

8

4 I 6 × I 2 + 2(I6 I6 − I 6 ) × I2 I2 012

1 8

− 18

541

I12

1 4

 

I 12

∗ − ∗ X1 X ∗2 (X ∗ 2 X 2 ) X 2 X 1 = 2 I8 I8

and C =

X 1 X1

∗ − ∗ − X1 X ∗2 (X ∗ 2 X2 ) X2 X1

= 2 I4 ×

1 −1 −1

1

.

It is clear then that rank C = 4 and hence (13.37) is satisfied since q1 = 8, n = 4. All main effects are estimable (which we knew already, of course). Let us now return to (13.36). Since C is not of full rank, we write a solution to (13.36) as 1 = C − Q E where C − is a generalized inverse of C. We also have for any estimable function c E 1 1 ) = c C − cσe2 var(c E The estimable functions here are, of course, those belonging to the various main effects. To illustrate this we consider again Example 13.6. Example 13.6 (Continued)

A generalized inverse for C is 

1      − C =     

2

0

0 1 2

0

0 1 2

0 1 2

           0

542


It is verified easily that ∗ − ∗ X 1 X∗2 (X ∗ 2 X2 ) X2 =

1 2

I8 I8

and hence 

Q = X1 −

1 2

1  −1   1 1  −1 I8 I8 y =  2 1  −1   1 −1

1 −1 −1 1 −1 1 1 −1

−1 1 1 −1 −1 1 1 −1

−1 1 −1 1 1 −1 1 −1

1 −1 1 −1 −1 1 −1 1

1 −1 −1 1 1 −1 −1 1

−1 1 1 −1 1 −1 −1 1

 −1 1  −1   1 y −1  1  −1  1

Finally, we obtain      1 − 1 = C Q =  E  4     We thus have

1 0 1 0 1 0 1 0

1 0 −1 0 −1 0 1 0

−1 0 1 0 −1 0 1 0

−1 0 −1 0 1 0 1 0

1 0 1 0 −1 0 −1 0

1 0 −1 0 1 0 −1 0

−1 0 1 0 1 0 −1 0

  y(0000) −1 0 y(0110)   −1  y(1010)   0 y(1100)   −1  y(0011)   0 y(0101)  −1  y(1001) y(1111) 0

  0 A   y(0 . . .) − y(1 . . .)   A1      0 B     0   y(. 0 . .) − y(. 1 . .)       B1    0  =  C  y(. . 0 .) − y(. . 1 .)  0       0 1   C     y(. . . 0) − y(. . . 1)     F0  0 1 F

and hence for the four estimable functions A = A1 − A0 , B = B1 − B0 , C = C1 − C0 , F = F1 − F0 , we obtain the estimators = A 1 − A 0 = y(1 . . .) − y(0 . . .) A = B 1 − B 0 = y(. 1 . .) − y(. 0 . .) B =C 1 − C 0 = y(. . 1 .) − y(. . 0 .) C = F 1 − F 0 = y(. . . 1) − y(. . . 0) F


with

= var(B) = var(C) = var(F ) = var(A)

This agrees, of course, with our earlier results.

543

1 2 2 σe

We noted that for resolution III designs the minimum number of distinct treatment combinations, that is, the number of treatment combinations in D0 , is 1 + q1 − n, and designs with the number of runs can be constructed. No such general statement, simply based on the rank condition (13.37), can be made for resolution IV designs because of the presence of nuisance parameters, for example, 2-factor interactions. It is possible, however, to give a lower bound (LB) of the number of runs n needed for a resolution IV design. For a s1n1 × s2n2 × · · · × sq q factorial with s1 < s2 < · · · < sq such a bound was derived by Margolin (1969) as

q ! (si − 1)ni − (sq − 2) (13.38) LB = sq i=1

(see also Seely and Birkes (1984) for a comment on its validity). Specifically for symmetrical factorials, that is, s n , we have the following lower bounds: s

LB

2 3 4 5

2n 3(2n − 1) 4(3n − 2) 5(4n − 3)

We can see that the design in Example 13.6 achieves this lower bound. But a quick look at Table 13.9 shows that only few of those designs achieve the lower bound, namely for s = 2, n = 8, = 4 with 16 runs and s = 4, n = 6, = 3 with 64 runs. 13.6.4 Foldover Designs Experiments are often initiated with resolution III designs because they are relatively small in size and provide some preliminary information about possibly active factors. However, the fact that main effects are not clear may make it sometimes difficult to interpret the results from such an experiment. One way to solve or partially solve this problem is to augment the original design with another fractional replicate such that the resulting design is a resolution IV design. For the 2n− III fraction this can be achieved by using the foldover principle or producing a foldover design (Box and Wilson, 1951). The idea behind a foldover design simply is to replace in each treatment combination of the original fraction all 0 levels by 1 levels and all 1 levels by 0 levels. We shall illustrate this principle in the following example.

544


Example 13.7 Consider the 25−2 III fraction given by I = ABCD = CDE = ABE

(13.39)

and with the design obtained by solving the equations x1 + x2 + x3 + x4

=0

x3 + x4 + x5 = 0 x1 + x2

(13.40)

+ x5 = 0

This yields the design points (00000) (11110) (00110) (11000)

(01101) (10011) (01011) (10101)

(11111) (00001) (11001) (00111)

(10010) (01100) (10100) (01010)

The foldover design then is

It is easy to see that the treatment combinations of the foldover design satisfy the equations x1 + x2 + x3 + x4

=0

x3 + x4 + x5 = 1 x1 + x2

+ x5 = 1

Alternatively, we could have written (13.39) more precisely as I = ABCD = −CDE = −ABE and then the identity relationship for the foldover fraction becomes I = ABCD = CDE = ABE so that the identity relationship for the combined design is simply I = ABCD that is, we have obtained a 25−1 IV fraction.

Within the practical context the augmentation of the original fraction by the foldover fraction results in a sequential design in that the two fractions can be


545

thought of as occurring in two different blocks. Formally this can be expressed by adding another factor to the original design and having that factor occur at the same level in the original fraction. Applying the foldover principle to the new design yields then the same results as described above. We illustrate this briefly with a continuation of Example 13.7 Example 13.7 (Continued) Denote the new (blocking) factor by F . We then have a 26−3 fraction defined by I = ABCD = −CDE = −ABE = F = ABCDF = −CDEF = −ABEF The foldover design is thus defined by I = ABCD = CDE = ABE = −F = −ABCDF = −CDEF = −ABEF The two fractions combined are then defined by I = ABCD = −CDEF = −ABEF

(13.41)

hence yielding a 26−2 IV fraction. Assuming no block treatment interaction, we see from (13.41) that with respect to the original factors A, . . . , E we obtain again a 25−1 IV design as before. From Example 13.7 we can easily deduce the general idea associated with the foldover principle: Starting with a resolution III design, we know that the identity relationship contains one or more words of length 3 and other words of length > 3. Or, alternatively, the defining equations [such as (13.40)] contain three or more unknowns. Using the foldover principle changes the right-hand sides of all equations with an odd number of unknowns, but does not change the right-hand sides of equations with an even number of unknowns. This implies that the identity relationship for the final design contains only interactions with an even word length, the smallest of which will be 4. The foldover principle is thus an easy way to achieve dealiasing of main effects and 2-factor interactions. Such dealiasing can also be achieved by constructing what we might call a partial foldover design. Rather than changing the level of each factor in each treatment combination, we may choose to change only the levels of a judiciously chosen subset of factors, in particular only one factor. This may have some merit from the practical point of view. It is, of course, clear that the foldover principle leads to a design with twice the number of design points as the original fraction. Especially for larger designs, that is, a large number of factors, this may be a disadvantage if not all main effects may need to be dealiased. Other techniques, such as optimality based techniques (see Wu and Hamada, 2000), may then be used instead by focusing on a particular model as suggested by the data from the original experiment.

546


13.6.5 Resolution V Designs For resolution V designs we wish to estimate the mean, main effects, and 2-factor interactions. Hence models (13.26) and (13.27) are appropriate. For purposes of our discussion we write the model for short as E[y] = X · E

(13.42)

X = (I, X1 , X2 )

(13.43)

where

and E = (M, E 1 , E 2 ) with q = 1 + q1 + q2 parameters. The NE for model (13.42) is (X X)E = X y and we know that because of the parameterization used rank(X X) < q We also know that the number of degrees of freedom associated with the mean and all main effects and 2-factor interactions included in E is 1+

n

(si − 1) +

(si − 1)(sj − 1) = ν(E)

i<j

i=1

say. For a design D to give rise to ν(E) degrees of freedom for SS(X), the sum of squares due to fitting the model (13.42), and hence to the estimation of the mean, all main effects and 2-factor interactions, we must have rank(X X) = ν(E) A design D with ν(E) runs can always be achieved, a result similar to that for resolution III designs. Such a design is obviously a saturated design. We illustrate this with the following example. Example 13.8 Consider n = 4 and s1 = s2 = s3 = s4 = 2. Then ν(E) = 11, and it can be shown that the following (nonorthogonal) FFD allows the estimation of all main effects and 2-factor interactions: (0000) (0011) (1100) (1111)

(1010) (1001) (0110) (0101)

(1110) (1101) (1011)

FRACTIONAL FACTORIALS AND COMBINATORIAL ARRAYS

547

13.7 FRACTIONAL FACTORIALS AND COMBINATORIAL ARRAYS The treatment combinations of a factorial design can conveniently be represented in the form of an array or a matrix with n rows representing the factors A1 , A2 , . . . , An and N columns representing the treatment combinations (runs). The elements of this array represent the levels of the factors, that is, if factor Ai has si levels, then the elements in the ith row take on values from the set Si = {0, 1, 2, . . . , si − 1} depending on which treatment combinations are part of the design. In this section we shall give an overview of several types of arrays and their relationship to certain types of fractional factorials. 13.7.1 Orthogonal Arrays Let us consider an orthogonal array of size N , with n constraints, s levels, strength d, and index λ, denoted by OA[N, n, s, d; λ] (see Appendix C). From our discussion above it follows that the rows represent the factors, the columns represent the runs, and each factor has s levels. We are considering here specifically the case N = s n− = λs d . Rao (1947b, 1950) has shown that for d = p + q − 1 we can estimate from such an array main effects and interactions involving up to p factors if interactions involving q(q > p) or more factors are assumed to be negligible. Of particular interest are the following cases: d

p

q

Resolution

2 3 4

1 1 2

2 3 3

III IV V

An extension of this result to asymmetrical factorial designs was given by n Chakravarti (1956). Suppose we consider the s1n1 × s2n2 × · · · × sg g factorial setting and we wish to construct a subset of size N from the complete asymmetrical factorial such that certain effects are estimable assuming that other effects (interactions) are negligible. The solution given by Chakravarti (1956) can be stated in the following theorem. Theorem 13.2 If orthogonal arrays OA[Ni , ni , si , di ; λi ] (i = 1, 2, . . . , g) exist, then the Kronecker product array of the g orthogonal arrays above yields a fractional factorial designs in g

g

i=1

i=1

N = Ni = λi sidi runs. If the ith orthogonal array is of strength di = pi + qi − 1 (with pi , qi as defined above), then in this Kronecker product array all main effects and g interactions involving up to r = i=1 ri (0 ≤ ri ≤ pi ; 0 < r ≤ pi ) factors are estimable, where ri factors are chosen from the ith group of factors at si levels (i = 1, 2, . . . , g).

548


As an illustration of this procedure we consider the following example. Example 13.9 (Chakravarti, 1956) Suppose we have three factors A1 , A2 , A3 with two levels each and four factors B1 , B2 , B3 , B4 with three levels each; that is, we have the 23 × 34 asymmetrical factorial. There exist resolution III designs for the individual systems with 4 and 9 runs, respectively, given by OA[22 , 3, 2, 2; 1] A1 A2 A3

1

2

3

4

0 0 0

1 1 0

1 0 1

0 1 1

and OA(32 , 4, 3, 2; 1)

B1 B2 B3 B4

1

2

3

4

5

6

7

8

9

0 0 0 0

0 1 2 1

0 2 1 2

1 1 1 0

1 2 0 1

1 0 2 2

2 2 2 0

2 0 1 1

2 1 0 2

which correspond to the identity relationships I = A1 A2 A3 and I = B1 B2 B3 = B2 B32 B4 = B1 B22 B4 = B1 B32 B42 respectively. The Kronecker product design is then obtained by combining every column of OA[22 , 3, 2, 2; 1] with every column of OA[32 , 4, 3, 2; 1] resulting in an array of n1 + n2 = 7 rows and N1 N2 = 4 · 9 = 36 columns. Since p1 = p2 = 1 and q1 = q2 = 2, it follows that in addition to the main effects only 2-factor interactions involving one factor Ai (i = 1, 2, 3) and one factor Bj (j = 1, 2, 3, 4), each with 2 d.f., are estimable assuming that all other interactions are negligible. This method can produce a large number of asymmetrical fractional factorials, but, as is the case with many Kronecker product designs, the number of runs becomes quite large. For this reason we mention only a few other special cases that lead to useful designs with the same properties as the design in Example 13.9: 1. For the 2n · s complete factorial (3 < n ≤ 7) combine each level of the factor with s levels with each column of the array OA[8, n, 2, 2; 2] to obtain a design with 8s runs. 2. For the 2n · s complete factorial (7 < n ≤ 11) combine each level of the factor with s levels with each column of the array OA[12, n, 2, 2; 3] to obtain a design with 12s runs.

549

BLOCKING IN FRACTIONAL FACTORIALS

3. For the 2n1 · 3n2 complete factorial with 3 < n1 ≤ 7 and 1 < n2 ≤ 4 obtain the Kronecker product design of OA[8, n1 , 2, 2; 2] and OA[32 , n2 , 3, 2; 1] with 72 runs. For a more general characterization of designs in the above classes as well as other factor combinations, see Chakravarti (1956). 13.7.2 Balanced Arrays Just as orthogonal arrays are related to fractional factorial designs, so are balanced arrays (see Appendix C). B-arrays of strength 2, 3, 4 represent resolution III, IV, V designs, respectively. As an illustration consider the following example. Example 13.10 BA(5, 4, 2, 2; (1, 1,  0 0 0 1 B= 0 1 0 1

2)) is given by  1 1 1 0 1 1  1 0 1 1 1 0

This represents a main effect plan for four factors in five runs.

However, contrary to fractional factorial designs in the form of orthogonal arrays, the designs from B-arrays are nonorthogonal, but as shown by Chakravarti (1956) and Srivastava (1965) the variance–covariance matrix for estimates of the i ) is the same parameters is balanced. For Example 13.10 this means that var(A i ) is the same for each i, and cov(A i , A j ) is the for each i = 1, 2, 3, 4, cov( µ, A same for each pair (i, j ) with i = j . In a series of studies Srivastava and Chopra (1971) and Chopra and Srivastava (1974, 1975) have provided methods of constructing B-arrays of strength 4 and a list of resolution V designs for 4 ≤ n ≤ 8 factors and a wide range of N runs. 13.8 BLOCKING IN FRACTIONAL FACTORIALS 13.8.1 General Idea As we have discussed, the purpose of FFDs is to reduce the number of experimental runs consonant with the objectives of the experiment. That may still leave us with a relatively large number of treatment combinations such that in the practical setting some form of blocking may be required. To achieve such blocked designs we simply need to combine the ideas of systems of confounding and fractional factorials. The concept can be described in general terms as follows. Let us consider an unreplicated s n− fractional factorial to be used in s m blocks. We then know the following: 1. We have s n− treatment combinations. 2. The available treatment combinations are determined by independent interaction components in the identity relationship.

550


3. We can estimate (s n− − 1)/(s − 1) functions of main effects and interaction components. 4. The estimable functions are determined by the identity relationship in the form of the alias structure. 5. We now have s m blocks of size 2n−−m . 6. We have s m − 1 d.f. among blocks. 7. We need to confound (s m − 1)/(s − 1) interaction components with blocks. 8. From the available estimable functions (as given by the alias structure) we need to select m functions to confound with blocks. 9. Then all their generalized interactions (and aliases) are also confounded with blocks. We shall illustrate these steps and some other ideas in the following sections with particular reference to the 2n and 3n systems. 13.8.2 Blocking in 2n− Designs Consider a 2n− FFD defined by the identity relationship (see Section 13.3.1) I = E α1 = E α2 = · · · = E α = E α1 +α 2 = · · · = E α1 +α2 +...+α

(13.44)

where E α1 , E α2 , . . . , E α are the independent interactions. We shall denote the 2n− − 1 = q estimable effects by E β 1 , E β 2 , . . . , E β q . The alias structure is obtained by multiplying each E β i (i = 1, 2, . . . , q) into (13.44). We now consider placing the treatment combinations specified by (13.44) into 2m blocks of size 2n−−m . To accomplish this we need to select m effects from among the E β i (i = 1, 2, . . . , q) to confound with blocks, say, without loss of generality, E β 1 , E β 2 , . . . , E β m . We know that then all their generalized interactions are also confounded with blocks, giving rise to the required 2m blocks. To generate the intrablock subgroup (IBS) we combine the equations for generating the fraction specified by (13.44) and the m equations for the specified system of confounding, that is, α11 x1 + α12 x2 + · · · + α1n xn = 0 α21 x1 + α22 x2 + · · · + α2n xn = 0 .. . α1 x1 + α2 x2 + · · · + αn xn = 0 β11 x1 + β12 x2 + · · · + β1n xn = 0 β21 x1 + β22 x2 + · · · + β2n xn = 0 .. . βm1 x1 + βm2 x2 + · · · + βmn xn = 0

(13.45)

551


Equations (13.45) show that the IBS constitutes a 2n−(+m) fraction, with the remaining blocks representing the other fractions obtained by changing successively the right-hand sides of the second set of equations in (13.45). In the practical context the interactions E β i (i = 1, 2, . . . , m) would have to be chosen carefully so as to avoid the confounding of “important” effects with blocks. We shall illustrate the preceding procedure with the following example. Example 13.11 Consider the 27−3 design in four blocks. We start with the resolution IV design defined by the identity relationship I = ABCD = CDEF = ABEF = BCEG = ADEG = BDFG = ACFG where the three independent interactions are underlined. The 15 estimable functions (d.f.) are accounted for by seven main effects (+ 3-factor and higher-order interactions), seven 2-factor interactions (+ higher-order interactions), and one 3-factor interaction (+ higher-order interactions). In order to set up blocks of size 4 we need to confound three 2-factor interactions (and their aliases), choosing, say, AB and AC as the two independent interactions. Thus, to generate the design we chose in (13.45) α 1 = (1 1 1 1 0 0 0)

α 2 = (0 0 1 1 1 1 0)

β 1 = (1 1 0 0 0 0 0)

β 2 = (1 0 1 0 0 0 0)

α 3 = (0 1 1 0 1 0 1)

An isomorphic design using SAS PROC FACTEX is given in Table 13.10.

13.8.3 Optimal Blocking As we have mentioned earlier, there may be various ways to set up the blocks for a particular fractional factorial, for example, by choosing different β j to use in (13.45). Such choices for the 2n− fraction in 2m blocks may be made by first inspecting the alias structure for the 2n− and then selecting m appropriate interactions to generate the system of confounding. We obviously need to be careful in this selection so that some of the resulting 2m − m − 1 generalized interactions do not end up being desirable effects. A more formal approach to this problem makes use of the fact that blocking in a 2n− can be considered as a special case of fractionation (Kempthorne, 1952; Lorenzen and Wincek, 1992) by introducing additional factors as blocking factors. More specifically, a 2n− fractional factorial in 2m blocks of size 2n−−m (m < n − ) can be considered formally as a 2(n+m)−(+m) (Chen and Cheng, 1999). The factors are divided into n treatment factors A1 , A2 , . . . , An and m blocking factors b1 , b2 , . . . , bm . The 2m “level” combinations of the blocking factors are used to divide the 2n− treatment combinations into 2m blocks. By treating this as a special case of a fractional factorial, the identity relationship will contain two types of words: treatment-defining words are those that contain only treatment

552


Table 13.10 27−3 Design in 4 Blocks options nodate pageno=1; proc factex; factors A B C D E F G; size design=16; model res=4; blocks size=4; examine design aliasing(3) confounding; output out=design blockname=block nvals=(1 2 3 4); title1 'TABLE 13.10'; title2 '2**(7-3) DESIGN IN 4 BLOCKS'; run; proc print data=design; run; The FACTEX Procedure Design Points Experiment Number

A

B

C

D

E

F

G

Block

--------------------------------------------1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

-1 -1 -1 -1 -1 -1 -1 -1 1 1 1 1 1 1 1 1

-1 -1 -1 -1 1 1 1 1 -1 -1 -1 -1 1 1 1 1

-1 -1 1 1 -1 -1 1 1 -1 -1 1 1 -1 -1 1 1

-1 1 -1 1 -1 1 -1 1 -1 1 -1 1 -1 1 -1 1

-1 1 1 -1 1 -1 -1 1 -1 1 1 -1 1 -1 -1 1

Factor Confounding Rules E = B*C*D F = A*C*D G = A*B*D

-1 1 1 -1 -1 1 1 -1 1 -1 -1 1 1 -1 -1 1

-1 1 -1 1 1 -1 1 -1 1 -1 1 -1 -1 1 -1 1

4 1 1 4 3 2 2 3 3 2 2 3 4 1 1 4

553


Table 13.10 (Continued ) Block Pseudofactor Confounding Rules [B1] = A*B*C*D [B2] = C*D Aliasing Structure A = B*D*G = B*E*F = C*D*F = C*E*G B = A*D*G = A*E*F = C*D*E = C*F*G C = A*D*F = A*E*G = B*D*E = B*F*G D = A*B*G = A*C*F = B*C*E = E*F*G E = A*B*F = A*C*G = B*C*D = D*F*G F = A*B*E = A*C*D = B*C*G = D*E*G G = A*B*D = A*C*E = B*C*F = D*E*F [B] = A*B = D*G = E*F A*C = D*F = E*G A*D = B*G = C*F [B] = A*E = B*F = C*G [B] = A*F = B*E = C*D A*G = B*D = C*E B*C = D*E = F*G A*B*C = A*D*E = A*F*G = B*D*F = B*E*G = C*D*G = C*E*F Obs 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

block 1 1 1 1 2 2 2 2 3 3 3 3 4 4 4 4

A

B

C

D

E

F

G

-1 -1 1 1 -1 -1 1 1 -1 -1 1 1 -1 -1 1 1

-1 -1 1 1 1 1 -1 -1 1 1 -1 -1 -1 -1 1 1

-1 1 -1 1 -1 1 -1 1 -1 1 -1 1 -1 1 -1 1

1 -1 1 -1 1 -1 1 -1 -1 1 -1 1 -1 1 -1 1

1 1 -1 -1 -1 -1 1 1 1 1 -1 -1 -1 -1 1 1

1 1 -1 -1 1 1 -1 -1 -1 -1 1 1 -1 -1 1 1

1 -1 1 -1 -1 1 -1 1 1 -1 1 -1 -1 1 -1 1

factors, and block-defining words are those that contain at least one blocking factor (Chen and Cheng, 1999). We shall illustrate this in the following example. Example 13.12 Consider the 27−3 in 22 blocks as a 2(7+2)−(3+2) fraction. We need three independent treatment-defining words, say A2 A3 A4 A5 , A1 A3 A4 A6 ,

554


and A1 A2 A4 A7 , and two independent block-defining words, say A1 A2 A3 A4 b1 and A3 A4 b2 . To obtain the desired design we thus need to solve the following equations: x2 + x3 + x4 + x5 x1 x1 + x2

+ x3 + x4

=0 + x6

+ x4

=0 + x7

x1 + x2 + x3 + x4

=0 + y1

x3 + x4

(13.46)

=0 + y2 = 0

where in (13.46) x1 , x2 , . . . , x7 refer to the treatment factors and y1 , y2 refer to the block factors, with xi = 0, 1, (i = 1, 2, . . . , 7), yj = 0, 1(j = 1, 2). The solution of Eqs. (13.46) leads to the design given in Table 13.10 if we use the following association between block numbers and block factor level combinations: Block No.

y1

y2

1 2 3 4

1 0 1 0

1 1 0 0

Inspection of the identity relationship in Example 13.12 illustrates some of the difficulties encountered in extending the notions of resolution and minimum aberration to situations involving block-defining words. The presence of a threeletter word, for example, A3 A4 b2 , would suggest that the design is a resolution III design, when in fact it is clearly a resolution IV design with regard to the treatment factors. Similar problems have been addressed and solutions proposed by, for example, Bisgaard (1994), Sitter, Chen, and Feder (1997), and Sun, Wu, and Chen (1997), but as Chen and Cheng (1999) point out, none of these solutions are satisfactory with regard to properly characterizing blocked designs and identifying optimal blocking schemes. The problem is that there are two different word length patterns, one for treatment-defining words and one for block-defining words. Suppose we consider only designs where no main effects are confounded either with other main effects or block effects. Following Sun, Wu, and Chen (1997), let Li,0 be the number of treatment-defining words containing i treatment letters, and let Li,1 be the number of block-defining words also containing i treatment letters and one or more block letters. We then have, for a given design, two types of word length patterns, one for treatment-defining words WLPt = (L3,0 , L4,0 , . . .)

(13.47)

555


and one for block-defining words WLPbt = (L2,1 , L3,1 , . . .)

(13.48)

Clearly, (13.47) and (13.48) have different meanings and are of different importance. An intuitive method to proceed would be to search for a design that has minimum aberration (see Section 13.3.4) with respect to both treatment and block factors. Zhang and Park (2000), however, have shown that no such design exists. Instead, in order to extend the concept of minimum aberration we somehow may need to combine (13.47) and (13.48). To this end Chen and Cheng (1999) propose the following ordering, based on the hierarchical assumption that lower-order effects are more important than higher-order effects and that effects of the same order are of equal importance: (L3,0 , L2,1 ), L4,0 , (L5,0 , L3,1 ), L6,0 , (L7,0 , L4,1 ) L8,0 , . . .

(13.49)

This means, for example, that if L3,0 > 0 some 2-factor interactions are aliased with main effects, and if L2,1 > 0 then some 2-factor interactions are confounded with blocks. In either case those 2-factor interactions cannot be estimated. Further, designs with L3,0 > 0 and/or L2, 1 > 0 are clearly less desirable than designs with L3,0 = L2,1 = 0 and L4,0 > 0. Based on such considerations, the criterion of minimum aberration for blocked fractional factorial designs might be defined by sequentially minimizing L3,0 , L2,1 , L4,0 , L5,0 , L3,1 , L6,0 , L7,0 , L4,1 , L8,0 , . . . . But Chen and Cheng (1999) show that even that is not quite satisfactory. Based on the notion of estimation capacity [for details see Cheng, Steinberg, and Sun, (1999)], they propose to minimize sequentially what they call the blocking word length pattern: (13.50) WLPb = Lb3 , Lb4 , . . . , Lbn+[n/2] where

Lbj =

 Lj,0     

      

for even j ≤ n

j (j + 1)/2 Lj,0 + L(j +1)/2,1

for odd j ≤ n

for n + 1 ≤ j ≤ n + [n/2] (13.51) and [n/2] is the largest integer less than or equal to n/2. As can be seen from (13.51) the word length for odd j ≤ n is a linear combination of the two associated word lengths enclosed in parentheses in (13.49). For example, Lb3 = 3L3,0 + L2,1 and Lb5 = 10 L5,0 + L3,1 . For detailed methods of identifying minimum aberration blocked fractional factorial designs we refer the reader to Chen and Cheng (1999). In Tables 13.11, 13.12, and 13.13 we give, in a different form, some of their results for 8-, 16-, 32-run minimum aberration 2n− designs in 2m blocks for n = 4, 5, . . . , 10; Lj −[n/2],1

556


Table 13.11 8-Run Minimum Aberration 2n− Designs Independent Defining Words n

m

Treatments

Blocks

4 4 5 6

1 1 2 3

1 2 1 1

ABCD ABCD ABD, ACE ABD, ACE , BCF

ABb 1 ABb 1 , ACb 2 BCb 1 ABCb 1

a Giving

WLPb a 2 6 8 15

1 1 1 3

0 0 2 4

only Lb3 , Lb4 , Lb5 .

Table 13.12 16-Run Minimum Aberration 2n− Designs Independent Defining Words

m

Treatments

Blocks

5 5 5 6 6 6 7

1 1 1 2 2 2 3

1 2 3 1 2 3 1

ABDb 1 ABb 1 , ACDb 2 ABb 1 , ACb 2 , ADb 3 ACDb 1 ABb 1 , ACDb 2 ABb 1 , ACb 2 , ADb 3 BCDb 1

7

3

2

7

3

3

8

4

1

8

4

2

8

4

3

9

5

1

9

5

2

10

6

1

10

6

2

ABCE ABCE ABCE ABCE , ABDF ABCE , ABDF ABCE , ABDF ABCE , ABDF , ACDG ABCE , ABDF , ACDG ABCE , ABDF , ACDG ABCE , ABDF , ACDG, BCDH ABCE , ABDF , ACDG, BCDH ABCE , ABDF , ACDG, BCDH ABE , ACF , ADG, BCDH , ABCDJ ABE , ACF , ADG, BCDH , ABCDJ ABE , ACF , BCG, ADH , BCDJ , ABCDK ABE , ACF , BCG, ADH , BCDJ , ABCDK

n

a Giving


WLPb a 0 2 10 0 3 15 0

1 1 1 3 3 3 7

2 4 0 4 8 0 7

9

7

0

21

7

0

4 14

0

ABb 1 , ACb 2

12 14

0

ABb 1 , ACb 2 , ADb 3

28 14

0

BCb 1 ,

16 14 84

BCb 1 , BDb 2

24 14 92

BDb 1

28 18 96

ABCb 1 , BDb 2

37 18 184

ABb 1 , ACb 2 ABb 1 , ACb 2 , ADb 3 ABb 1

557


Table 13.13 32-Run Minimum Aberration 2n− Designs Independent Defining Words n

Treatments

m

6 6 6 6

1 1 1 1

1 2 3 4

ABCDEF ABCF ABCDEF ABCDEF

7 7 7

2 2 2

1 2 3

ABCF , ABDEG ABCF , ABDG ABCF , ABDEG

7

2

4

ABCF , ADEG

8

3

1

8

3

2

8

3

3

8

3

4

9

4

1

9

4

2

9

4

3

9

4

4

10

5

1

10

5

2

10

5

3

10

5

4

ABCF , ABDG, ACDEH ABCF , ABDG, ACDEH ABCF , ABDG, ACEH ABCF , ABDG, ACEH ABCF , ABDG, ABEH , ACDEJ ABCF , ABDG, ACEH , ADEJ ABCF , ABDG, ACEH , ADEJ ABCF , ABDG, ACEH , ADEJ ABCF , ABDG, ACEH , ADEJ , ABCDEK ABCF , ABDG, ACEH , ADEJ , ABCDEK ABCF , ABDG, ACDH , ABEJ , ACEK ABCF , ABDG, ACEH , ADEJ , ABCDEK

a Giving


Blocks ABCb 1 ABDb 1 , ACEb 2 ABb 1 , CDb 2 , ACEb 3 ABb 1 , ACb 2 , ADb 3 , AEb 4 ACDb 1 ACDb 1 , ABEb 2 ACb 1 , ABDb 2 , ABEb 3 ABb 1 , ACb 2 ADb 3 , AEb 4 ABEb 1

WLPb a 0 0 3 15

0 1 0 0

2 4 8 0

0 1 22 0 3 7 5 1 32 21 2 0 0 3 43

ABEb 1 , BCDEb 2

1 3 50

ABb 1 , ACDb 2 , AEb 3 ABb 1 , ACb 2 , ADb 3 , AEb 4 BCDEb 1

7 5 18 28 5 0

BCb 1 , BDEb 2

2 9 14

ABb 1 , ACDb 2 , AEb 3 ABb 1 , ACb 2 , DEb 3 , AEb 4 ACDb 1

9 9 27 36 9 0

ABb 1 , ACDb 2

3 15 20

ABb 1 , ACb 2 , ADEb 3

12 16 36

ABb 1 , ACb 2 ADb 3 , AEb 4

45 15 0

0 6 84

0 15 10

558


= 1, 2 . . . , 6; m = 1, 2, 3, 4. For possible combinations of n, , and m we give the independent words for the identity relationship from which the design can be obtained by solving a set of equations as illustrated in Example 13.12. We also list the first three components, Lb3 , Lb4 , and Lb5 , of each associated WLPb . In Tables 13.11, 13.12, and 13.13 we denote the treatment factors by A, B, C, D, E, F , G, H , J , K and the blocking factors by b1 , b2 , b3 , b4 . 13.9 ANALYSIS OF UNREPLICATED FACTORIALS One reason for using fractional factorial designs is to reduce the number of runs in the experiments. But even a highly fractional design for a large number of factors may involve quite a large number of runs. In that case we may be able to afford only a few or no replications of the treatment combinations included in the design. As we know replications are important for estimating the variance of the estimable effects and interactions. For unreplicated 2n factorials we have described how negligible interactions can be used to estimate σe2 (see Section 7.6.6). This method is, of course, applicable also for the general s n factorial and mixed factorials. The problem becomes more complicated for fractional factorials where often only main effects and low-order interactions are estimable. In this case we have to fall back on the assumption of effect sparsity, which says that typically only a few effects and/or interactions are real. We shall mention here briefly the essential ideas of two methods that are helpful with the analysis of data and the interpretation of the results from unreplicated fractional factorials. 13.9.1 Half-Normal Plots Let us consider a 2n− fraction and suppose that the effects and interactions E α1 , E α2 , . . ., E αm (m = 2n− − 1) are estimable (these are, of course, aliased with αi | = other interactions which we assume to be negligible). Noting that | E αi αi | E1 − E0 | can be viewed as the range of a pair of (approximately) norαi | are themselves halfmally distributed random variables suggests that the | E normally distributed, that is, follow the distribution function f (x) = [2/(π σe2 )]1/2 exp[−x 2 /(2 σe2 )] =0

for x ≥ 0 for x < 0

Its cdf can be linearized by using the upper half of normal probability paper, relabeling the P axis for P ≥ .5 as P = 2P − 1. Based on this idea Daniel (1959) proposed a procedure that has come to be known as half-normal plot. It consists of plotting the absolute values of the m estimates, approximating their P values by P = (i − 12 )/m for i = 1, 2, . . . , m. If, as assumed, most of the estimates are, in fact, estimating zero, then they should fall near a straight line through the origin. The 68% point (obtained from P = 2 × .84 − 1 = .68)

559

ANALYSIS OF UNREPLICATED FACTORIALS

would provide an approximate value of the standard error, say SE0 , of the effects αi ) = plotted, that is, an estimate of σe2 since SE(E σe /(2n−−2 )1/2 . Furthermore, to the extent that some of the largest estimates deviate from the straight line is an indication that the corresponding effects are different from zero. The above describes in essence the procedure, but we need to elaborate on two points: (1) How do we judge an estimate to be significantly different from zero? and (2) should we reestimate σe2 in light of some significant estimates? 1. The significance of effects is judged by whether they fall outside certain “guardrails” (Daniel, 1959). The α-level guardrails are constructed by drawing a line through the α-level critical values for the largest standardized absolute effects, say th (h = m, m − 1, . . . , m − K), standardized by the initial estimate of σe . Zahn (1975) provided the critical values for m = 15, 31, 63, 127 and α = .05, .20, .40. In Table 13.14 we give the critical values for m = 15, 31 and α = .05, .20, .40 for some th . 2. Based upon the test performed by using the procedure in point 1 above, we exclude not only the significant effects at α = .05 but also “borderline” effects at α = .20 or even α = .40, say there are Kα such effects altogether, and replot the m − Kα remaining effects. The value at P = .68 provides αi ), say SEf , and from it we obtain the final then the final value for SE(E estimate of σe [for a more refined explanation see Zahn (1975)]. The half-normal plot procedure is illustrated in the following example. Example 13.13 The following data are taken from Example 10.4 in Davies (1956), which describes a 25−1 factorial investigating the quality of a basic dyestuff. The factors are A (temperature), B (quality of starting material), C Table 13.14

Guardrail Critical Values

m = 15:

h

α = .05

α = .20

15

3.230

2.470

2.066

14

2.840

2.177

1.827

13

2.427

1.866

1.574

12

2.065

1.533

1.298

h

α = .05

α = .20

31

3.351

2.730

2.372

30

3.173

2.586

2.247

29

2.992

2.439

2.121

28

2.807

2.288

1.991

27

2.615

2.133

1.857

m = 31:

α = .40

α = .40

560


(reduction pressure), D (pressure), and E (vacuum leak). The 12 -fraction is based on I = −ABCDE. The treatment combinations and observations Y are given in Table 13.15. Pertinent computed values to draw the half-normal plot are given in Table 13.16, and the half-normal plot is given in Figure 13.1. = 8.31, the 11th-order statistic, is From Table 13.16 we see that SE0 = BC αi ). The closest to the .68 percentile and hence is the initial estimate of SE(E α i | /8.31. The guardrails for standardized effects are then obtained as ti =| E α = .05, .20, and .40 are constructed by drawing lines through the corresponding Table 13.15 25−1 Fractional Factorial Treatment 00000 10001 01001 11000 00101 10100 01100 11101

Y

Treatment

201.5 178.0 183.5 176.0 188.5 178.5 174.5 196.5

00011 10010 01010 11011 00110 10111 01111 11110

Y 255.5 240.5 208.5 244.0 274.0 257.5 256.0 274.5

Table 13.16 Calculations for Half-Normal Plot of Figure 13.1 Effect

αi E

Order(i)

(i − .5)/15

ti

DE

− .0625

1

.0333

.00752

A

.4375

2

.1000

.05263

AE

−2.3125

3

.1667

.27820

AC

3.0625

4

.2333

.36842

BD

−3.5625

5

.3000

.42857

E

3.9375

6

.3667

.47368

CE

−4.6875

7

.4333

.56391

AD

5.1875

8

.5000

.62406

B

−7.5625

9

.5667

.90977

BE

7.6875

10

.6333

.92481

BC

8.3125

11

.7000

1.00000

C

14.0625

12

.7667

1.69173

CD

14.3125

13

.8333

1.72180

AB

16.6875

14

.9000

2.00752

D

66.6875

15

.9667

8.02256

561

ANALYSIS OF UNREPLICATED FACTORIALS 9

Variable standardized abs(effect) critical values_0.05 critical values_0.2 critical values_0.4

D

Standardized Absolute Effect

8 7 6 5 4 3 2

C CD

1 DE A

E CE AD AE AC BD

AB

B BE BC

0 0

2

4

6

8 10 11 12 13 14 15 16 Order

Figure 13.1

Half-normal plot with guardrails.

9 8

Absolute Effect

7 6 AD 5 4 3 2 1 0 0.045

0.136

0.227

0.318

0.409

0.500

0.591

0.682

0.773

0.864

0.954

Percentile

Figure 13.2

Half-normal plot for nonsignificant effects.

critical values for t15 , t14 , t13 , and t12 . From Figure 13.1 we see that for α = .05 only D is significant, for α = .20 D and C are significant, and for α = .40 D, C, CD, and AB are significant. Using the latter result, we obtain the final estimate of σe from the .68 percentile of the half-normal plot for the 11 smallest effects in = 5.188 = SE(E αi ) = Figure 13.2. From it we see that AD σe /(25−1−2 )1/2 = SEf , and hence σe = 2 × 5.188 = 10.376.

562


13.9.2 Bar Chart This procedure, proposed by Lenth (1989), is based on a simple formula for the standard error of the estimated effects. The effects are plotted and the usual t procedures are used to judge significance and interpret the results. Let αi | s0 = 1.5 × median | E and define the pseudostandard error (PSE) of the estimated effect by αi | PSE = 1.5 × r-median | E

(13.52)

αi | < 2.5s0 . PSE (correwhere r-median is the restricted median over all | E sponding to SEf in Section 13.9.1) in (13.52) is an estimate of σe /(2n−−2 )1/2 . It can be used to define two useful quantities that in turn can be used to assess the significance of effects. Let ME = t.975,d × PSE

(13.53)

αi , where t.975,d is the .975th quantile of the t denote the margin of error for E distribution with d = m/3 d.f. (m = 2n− − 1). An approximate 95% confidence αi ± ME. A simultaneous margin of error interval for E αi is then given by E (SME) is defined: SME = tγ ,d × PSE

(13.54)

where γ = (1 + .951/m )/2. For some values of m we have m

t.975,d

tγ ,d

7

3.76

9.01

15

2.57

5.22

31

2.22

4.22

63

2.08

3.91

The procedure proposed by Lenth (1989) then is as follows: 1. Construct a bar chart of the signed effects. 2. Add reference lines at ± ME [from (13.53)] and ± SME [from (13.54)]. αi extends beyond the SME lines, 3. Judge an effect as clearly active if E α i does not extend beyond the ME lines, and of debatable inactive if E αi falls between the ME and SME lines. significance if E

563

ANALYSIS OF UNREPLICATED FACTORIALS 70 60 50

Effect Size

40

39.05

30 20

19.23

10 0 –10 –19.23

–20 A

B

C

D

E

AB

AC

AD

AE

BC

BD

BE

CD

CE

DE

Effect

Figure 13.3

Bar chart for Example 13.14.

We illustrate this procedure in the following example. Example 13.14 Consider the data from Example 13.13. We obtain s0 = 1.5 × 5.1875 = 7.78, PSE = 1.5 × 4.9875 = 7.48, ME = 19.23, SME = 39.05. The bar chart is given in Figure 13.3. From it we see that only D is significant at α = .05. This agrees with the result in Example 13.13.

13.9.3 Extension to Nonorthogonal Design The previous analysis procedures are applicable for orthogonal designs. There are, however, situations where we may encounter nonorthogonal designs, either by construction (see Section 14.4) or as a consequence of some mishap in the experiment. For example, we may have some replicates for one or a few treatment combinations. Or we may have missing values. Both cases lead to nonorthogonal designs, complicating the analysis and interpretation of results. A data-driven method for nonorthogonal designs was proposed by Kunert (1997), extending the method of Daniel (1959) as modified by Zahn (1975) and the method of Lenth (1989). For details we refer the reader to the original article by Kunert (1997).

C H A P T E R 14

Main Effect Plans

14.1 INTRODUCTION Main effect plans (MEP) or resolution III designs are commonly used in exploratory studies when a large number of factors need to be considered or screened. The idea is to try to detect factors that exhibit “large” main effects and discard factors with no noticeable effects from further study. The crucial assumption here is that all interactions are negligible, including 2-factor interactions. This may not always be realistic, but as a first approximation this is nevertheless a very valuable method. In some cases it may be helpful to know with which 2-factor interactions each main effect is aliased. We shall make some remarks about that later.

14.2 ORTHOGONAL RESOLUTION III DESIGNS FOR SYMMETRICAL FACTORIALS Among the resolution III designs those that admit uncorrelated estimates of main effects are of major importance. Not only are they easy to construct and analyze, but the results from such experiments are also easy to interpret. Because of their special feature, these designs are referred to as orthogonal resolution III designs or orthogonal main effect plans (OMEP). We shall now describe some methods of constructing OMEPs. 14.2.1 Fisher Plans In discussing systems of confounding (see Section 11.4) for an s n factorial in blocks of size s with n = (s − 1)/(s − 1) we have already provided a method of constructing a 1/s n− fraction of an s n factorial. Such a FFD is a saturated Design and Analysis of Experiments. Volume 2: Advanced Experimental Design By Klaus Hinkelmann and Oscar Kempthorne ISBN 0-471-55177-5 Copyright  2005 John Wiley & Sons, Inc.

564

ORTHOGONAL RESOLUTION III DESIGNS FOR SYMMETRICAL FACTORIALS

565

n−(n−) OMEP, designated by sIII , indicating that this is a 1/s n− fraction of an n n−(n−) = s treatment combinations (runs, trials) and s factorial resulting in s m is of resolution III. Here s = p and p is a prime. The method of construction is based on results given by Fisher (1942, 1945) on how to obtain the IBSG for an s n factorial in blocks of size s . The IBSG is then the OMEP. The method of construction guarantees that each level of each factor occurs the same number of times, namely s −1 , and each level combination for any two factors occurs the same number of times, namely s −2 (see Section 11.3). The latter is a sufficient condition for the estimates of the main effects to be uncorrelated. To see this, consider model (13.28) with X1 having the following properties:

1. Each column has s −1 unity elements, the remaining s −1 (s − 1) elements are zero. 2. The inner product of any two columns belonging to the same factor [i.e., columns belonging to (Ai )δ and (Ai )γ , δ, γ = 0, 1, . . . , s − 1; δ = γ ] is equal to zero. 3. The inner product of any two columns belonging to different factors [i.e., columns belonging to (Ai )δ and (Aj )γ , i, j = 1, 2, . . ., n; i = j ] is equal to s −2 . Considering (13.28) as a two-part linear model and using the result of Section I.4.8 we can write the RNE for E 1 as 1 1 X1 I − II X1 E 1 = X1 I − II y s s

(14.1)

The coefficient matrix in (14.1) can be written alternatively as

1 1 1 X 1 X1 I − II I − II X1 = X s s

(14.2)

with 1 = I − 1 II X1 X s 1 is obtained from X1 by subtracting from each element the correspondthat is, X ing column mean. Because of property 1 each column mean is 1/s. Because of properties 2 and 3, it follows then that 1 X 1 = diag [M, M, . . . , M] X n

566

MAIN EFFECT PLANS

with M an s × s matrix of the form M = s −2 (sI − II ) Using the result from Theorem 1.1 we then have −

1 1 X = diag(M − , M − , . . . , M − ) X

(14.3)

with M− =

1 I s −1

−

1 1 X is the variance–covariance matrix for estimable functions Since X associated with the various main effects, it follows then from (14.3) that

estimable functions associated with different main effects, such as δ cδ (Ai )δ

and δ dδ (Aj )δ for the two factors Ai and Aj , are uncorrelated. The same, of course, is true for the estimates of orthogonal contrasts associated with the main

i )δ and

effects of the same factor, such as δ cδ (A δ dδ (Ai )δ with δ cδ dδ = 0. The estimates of the estimable functions are obtained from (14.1) as −

1 X 1 y

1 = X 1 X E Examples of OMEPs that can be obtained by this method are given in Table 14.1. In this table n represents the maximum number of factors that can be accommodated in N = s runs resulting in a saturated OMEP. It should be clear how this method can also be used to construct OMEPs for n∗ < n = (s − 1)/(s − 1) factors in N runs. All we need to do is to construct the plan for n factors and then delete n − n∗ factors from that plan. The resulting design is obviously an OMEP, but it is no longer a saturated OMEP. Instead it provides (n − n∗ )(s − 1) d.f. for error. Table 14.1 Orthogonal Main Effect Plans (Fisher’s Method) sn

N = s

27 215 231 34 313 340 45 421 56

8 16 32 9 27 81 16 64 25


567

14.2.2 Collapsing Factor Levels The condition of equal frequencies for all level combinations for any pair of factors in an OMEP can be relaxed to that of proportional frequencies. This was shown first by Plackett (1946) and later by Addelman and Kempthorne (1961) for asymmetrical factorials (see Theorem 14.1). This result leads to the construction of additional OMEPs by using what is called collapsing of factor levels. For example, we may collapse the levels of a three-level factor to those of a two-level factor by setting up the correspondence Three-Level Factor

Two-Level Factor −→ −→ −→

0 1 2

0 1 0

Using this correspondence we can, for example, use the 34−2 III and obtain an 4 OMEP for the 2 factorial with nine runs. 14.2.3 Alias Structure Even though for resolution III designs we make the assumption that all interactions are negligible, it may in certain situations be useful to know with which 2-factor interactions the main effects are confounded. The reason, of course, is that among interactions the 2-factor interactions are least likely to be negligible and hence to know the aliasing between main effects and 2-factor interactions may be helpful for diagnostic purposes (see also Section 13.6.4). We shall illustrate this later. The general method of investigating the alias structure is, of course, to make use of models (13.28) and (13.33). That is, we estimate the main effects from a given design using (13.31) and then consider the expected value of the estimates under model (13.33). To illustrate this we have from (13.31), with E ∗ 1 = (M, E 1 ),

∗1 = (X∗ X∗ )− X∗ y E 1 1 1 and, using (13.33) ∗

1 ] = X∗ (X∗ X∗ )− X∗ (X∗ E ∗ + X2 E 2 ) E[X∗1 E 1 1 1 1 1 1 ∗ − ∗ = X∗1 E ∗1 + X∗1 (X ∗ 1 X1 ) X1 X2 E 2

or, for an estimable function c E ∗1 , ∗

1 ] = c E ∗ + c (X ∗ X∗ )− X ∗ X2 E 2 E[c E 1 1 1 1

(14.4)

568

MAIN EFFECT PLANS

Expressions for this alias structure can be stated explicitly (see below) for certain OMEPs. In other cases computer programs may be useful to obtain expressions for (14.4). Since the OMEP s n−(n−) with n = (s − 1)/(s − 1) is equivalent to the IBSG of an s n experiment in blocks of size s , it is obvious that the interactions that are confounded with blocks are also the interactions included in the identity relationship. Altogether there are (s n− − 1)/(s − 1) interactions in the identity relationship, n − independent interactions plus all their GIs. For our purpose here we are concerned only with knowing the 3-factor interactions contained in the identity relationship. Some 3-factor interactions are directly identified from (11.27), namely those E γ j for which the partition α j + corresponds to a 2-factor interaction in the s system. Others may be obtained as GIs among the n − γ independent interactions (11.27). Now if, for example, Fi 1 Fiδ2 Fi 3 is a 3-factor interaction in the identity relationship for some i1 , i2 , i3 with 1 ≤ i1 , i2 , i3 ≤ n and i1 , i2 , i3 different and some δ, γ with 1 ≤ δ, γ ≤ s − 1, then Fi1 will be conγ γ founded with Fiδ2 Fi3 (or an appropriate power of Fiδ2 Fi3 to bring it into standard form), and so forth. An alternative method, using existing computer programs, of obtaining the relevant alias structure is given below. We now illustrate our discussion with the following example. Example 14.1 Suppose s = 3, = 2, and hence n = 4. The 34−2 III was given as the IBSG in Example 11.2. From Example 11.4 it follows that the identity relationship is given by I = F1 F2 F32 = F1 F22 F42 = F1 F3 F4 = F2 F3 F42

(14.5)

or, more precisely, I = (F1 F2 F32 )0 = (F1 F22 F42 )0 = (F1 F3 F4 )0 = (F2 F3 F42 )0

(14.6)

where F1 F2 F32 and F1 F22 F42 are the independent interactions obtained from (11.27) and F1 F3 F4 , F2 F3 F42 are their GIs. It follows then from (14.5) that the relevant aliasing for F1 , for example, is given by F1 = F2 F32 = F2 F4 = F3 F4 If the 2-factor interactions were indeed not negligible, then instead of estimating, for example, (F1 )2 − (F1 )0 we would actually estimate [see also (14.4)] [(F1 )2 − (F1 )0 ] + [(F2 F32 )1 − (F2 F32 )0 ] + [(F2 F4 )2 − (F2 F4 )0 ] + [(F3 F4 )1 − (F3 F4 )0 ]

(14.7)

making use of (14.6). Clearly, if all 2-factor interactions were nonnegligible using the OMEP would be useless. But suppose the investigator is fairly sure

569


that all 2-factor interactions except possibly F2 × F3 are negligible. Then we would have to be concerned about the confounding of F1 and F2 F32 and F4 and F2 F3 . If then an estimated contrast belonging to F1 were “large,” we may suspect that this is partly due to a corresponding contrast belonging to F2 F32 , and so forth. Of course, we have no way of knowing, but this may suggest at least further investigation, for example, by augmenting the original design to achieve the desired dealiasing. It is in this sense that knowledge of the alias structure can be helpful as a diagnostic tool. The form of the estimable function (14.7) can also be obtained by using the E-option in SAS PROC GLM. The (fictitious) data and the ensuing analysis are given in Table 14.2. Concerning the SAS input and output in Table 14.2 we make the following comments: 1. The factors are now labeled A, B, C, D. 2. The model includes all 2-factor interactions similar to the model for a resolution IV design. 3. The option e1 lists all type I estimable functions (see comment 4) and produces the type I sums of squares (see I.4.11.1). 4. Using a model similar to (13.23) (see also Section 7.2.1), we see that by putting L2 = 1 and all other Li = 0(i = 3, 5, 6, 8, 9, 11, 12) the following function is estimable: α0 − α2 + 13 [(αβ)00 + (αβ)01 + (αβ)02 − (αβ)20 − (αβ)21 − (αβ)22 ] +

1 3 [(αγ )00

+ (αγ )01 + (αγ )02 − (αγ )20 − (αγ )21 − (αγ )22 ]

+

1 3 [(αδ)00

+ (αδ)01 + (αδ)02 − (αδ)20 − (αδ)21 − (αδ)22 ]

+

1 3 [(βγ )00

− (βγ )02 − (βγ )10 + (βγ )11 − (βγ )21 + (βγ )22 ]

+

1 3 [(βδ)00

− (βδ)02 − (βδ)11 + (βδ)12 − (βδ)20 + (βδ)21 ]

+

1 3 [(γ δ)00

− (γ δ)01 − (γ δ)10 + (γ δ)12 + (γ δ)21 − (γ δ)22 ]

(14.8)

Defining the interaction terms such that, for example, 2

(αβ)ij = 0 for

i = 0, 1, 2

j =0

then in (14.8) all terms in (αβ)ij , (αγ )ik , (αδ)i drop out. Concerning the remaining interaction terms, we make use of the two equivalent parameterizations, for example, B × C = BC + BC 2

570

MAIN EFFECT PLANS

Table 14.2 Orthogonal Main Effect Plan 34−2 options nodate pageno=1; data OMEP; input A B C D Y @@; datalines; 0 0 0 0 5

1 1 2 0 7

2 2 1 0 6

0 1 1 2 8

1 0 1 1 4

0 2 2 1 5

1 2 0 2 8

2 0 2 2 9

2 1 0 1 7

; run; proc print data=OMEP; title1 'TABLE 14.2'; title2 'ORTHOGONAL MAIN EFFECT PLAN 3**(4-2)'; run; proc glm data=OMEP; class A B C D; model y=A B C D A*B A*C A*D B*C B*D C*D/e1; title3 'ANALYSIS AND ESTIMABLE FUNCTIONS'; run; Obs

A

B

C

D

Y

1

0

0

0

0

5

2

1

1

2

0

7

3

2

2

1

0

6

4

0

1

1

2

8

5

1

0

1

1

4

6

0

2

2

1

5

7

1

2

0

2

8

8

2

0

2

2

9

9

2

1

0

1

7

The GLM Procedure

ANALYSIS AND ESTIMABLE FUNCTIONS

Class Level Information

Class

Levels

Values

A

3

0 1 2

B

3

0 1 2

C

3

0 1 2

D

3

0 1 2


571

Table 14.2 (Continued ) Number of observations 9 Type I Estimable Functions - - - - - - - - - - - - - - - - - - Coefficients - - - - - - - - - - - - - - - - - Effect

A

B

C

Intercept

0

0

0

A

0

L2

0

0

A

1

L3

0

0

A

2

-L2-L3

0

0

B

0

0

L5

0

B

1

0

L6

0

B

2

0

-L5-L6

0

C

0

0

0

L8

C

1

0

0

L9

C

2

0

0

-L8-L9

D

0

0

0

0

D

1

0

0

0

D

2

0

0

0

A*B

0 0

0.3333*L2

0.3333*L5

0.3333*L8

A*B

0 1

0.3333*L2

0.3333*L6

0.3333*L9

A*B

0 2

0.3333*L2

-0.3333*L5-0.3333*L6

-0.3333*L8-0.3333*L9

A*B

1 0

0.3333*L3

0.3333*L5

0.3333*L9

A*B

1 1

0.3333*L3

0.3333*L6

-0.3333*L8-0.3333*L9

A*B

1 2

0.3333*L3

-0.3333*L5-0.3333*L6

0.3333*L8

A*B

2 0

-0.3333*L2-0.3333*L3

0.3333*L5

-0.3333*L8-0.3333*L9

A*B

2 1

-0.3333*L2-0.3333*L3

0.3333*L6

0.3333*L8

A*B

2 2

-0.3333*L2-0.3333*L3

-0.3333*L5-0.3333*L6

0.3333*L9

A*C

0 0

0.3333*L2

0.3333*L5

0.3333*L8

A*C

0 1

0.3333*L2

0.3333*L6

0.3333*L9

A*C

0 2

0.3333*L2

-0.3333*L5-0.3333*L6

-0.3333*L8-0.3333*L9

A*C

1 0

0.3333*L3

-0.3333*L5-0.3333*L6

0.3333*L8

A*C

1 1

0.3333*L3

0.3333*L5

0.3333*L9

A*C

1 2

0.3333*L3

0.3333*L6

-0.3333*L8-0.3333*L9

A*C

2 0

-0.3333*L2-0.3333*L3

0.3333*L6

0.3333*L8

A*C

2 1

-0.3333*L2-0.3333*L3

-0.3333*L5-0.3333*L6

0.3333*L9

A*C

2 2

-0.3333*L2-0.3333*L3

0.3333*L5

-0.3333*L8-0.3333*L9

A*D

0 0

0.3333*L2

0.3333*L5

0.3333*L8

A*D

0 1

0.3333*L2

-0.3333*L5-0.3333*L6

-0.3333*L8-0.3333*L9

A*D

0 2

0.3333*L2

0.3333*L6

0.3333*L9

A*D

1 0

0.3333*L3

0.3333*L6

-0.3333*L8-0.3333*L9

A*D

1 1

0.3333*L3

0.3333*L5

0.3333*L9

572

MAIN EFFECT PLANS

Table 14.2 (Continued ) Type I Estimable Functions

- - - - - - - - - - - - - - - - - Coefficients - - - - - - - - - - - - - - - - Effect

D

A*B

A*C

A*D

B*C

B*D

C*D

Intercept

0

0

0

0

0

0

0

A

0

0

0

0

0

0

0

0

A

1

0

0

0

0

0

0

0

A

2

0

0

0

0

0

0

0

B

0

0

0

0

0

0

0

0

B

1

0

0

0

0

0

0

0

B

2

0

0

0

0

0

0

0

C

0

0

0

0

0

0

0

0

C

1

0

0

0

0

0

0

0

C

2

0

0

0

0

0

0

0

D

0

L11

0

0

0

0

0

0

D

1

L12

0

0

0

0

0

0

D

2

-L11-L12

0

0

0

0

0

0

A*B

0 0

0.3333*L11

0

0

0

0

0

0

A*B

0 1

-0.3333*L11-0.3333*L12

0

0

0

0

0

0

A*B

0 2

0.3333*L12

0

0

0

0

0

0

A*B

1 0

0.3333*L12

0

0

0

0

0

0

A*B

1 1

0.3333*L11

0

0

0

0

0

0

A*B

1 2

-0.3333*L11-0.3333*L12

0

0

0

0

0

0

A*B

2 0

-0.3333*L11-0.3333*L12

0

0

0

0

0

0

A*B

2 1

0.3333*L12

0

0

0

0

0

0

A*B

2 2

0.3333*L11

0

0

0

0

0

0

A*C

0 0

0.3333*L11

0

0

0

0

0

0

A*C

0 1

-0.3333*L11-0.3333*L12

0

0

0

0

0

0

A*C

0 2

0.3333*L12

0

0

0

0

0

0

A*C

1 0

-0.3333*L11-0.3333*L12

0

0

0

0

0

0

A*C

1 1

0.3333*L12

0

0

0

0

0

0

A*C

1 2

0.3333*L11

0

0

0

0

0

0

A*C

2 0

0.3333*L12

0

0

0

0

0

0

A*C

2 1

0.3333*L11

0

0

0

0

0

0

A*C

2 2

-0.3333*L11-0.3333*L12

0

0

0

0

0

0

A*D

0 0

0.3333*L11

0

0

0

0

0

0

A*D

0 1

0.3333*L12

0

0

0

0

0

0

A*D

0 2

-0.3333*L11-0.3333*L12

0

0

0

0

0

0

A*D

1 0

0.3333*L11

0

0

0

0

0

0

A*D

1 1

0.3333*L12

0

0

0

0

0

0

573


Table 14.2 (Continued ) Type I Estimable Functions - - - - - - - - - - - - - - - - - - Coefficients - - - - - - - - - - - - - - - - - Effect

A

B

C 0.3333*L8

A*D

1 2

0.3333*L3

-0.3333*L5-0.3333*L6

A*D

2 0

-0.3333*L2-0.3333*L3

-0.3333*L5-0.3333*L6

0.3333*L9

A*D

2 1

-0.3333*L2-0.3333*L3

0.3333*L6

0.3333*L8

A*D

2 2

-0.3333*L2-0.3333*L3

0.3333*L5

-0.3333*L8-0.3333*L9

B*C

0 0

0.3333*L2

0.3333*L5

0.3333*L8

B*C

0 1

0.3333*L3

0.3333*L5

0.3333*L9

B*C

0 2

-0.3333*L2-0.3333*L3

0.3333*L5

-0.3333*L8-0.3333*L9

B*C

1 0

-0.3333*L2-0.3333*L3

0.3333*L6

0.3333*L8

B*C

1 1

0.3333*L2

0.3333*L6

0.3333*L9

B*C

1 2

0.3333*L3

0.3333*L6

-0.3333*L8-0.3333*L9

B*C

2 0

0.3333*L3

-0.3333*L5-0.3333*L6

0.3333*L8

B*C

2 1

-0.3333*L2-0.3333*L3

-0.3333*L5-0.3333*L6

0.3333*L9

B*C

2 2

0.3333*L2

-0.3333*L5-0.3333*L6

-0.3333*L8-0.3333*L9

B*D

0 0

0.3333*L2

0.3333*L5

0.3333*L8

B*D

0 1

0.3333*L3

0.3333*L5

0.3333*L9

B*D

0 2

-0.3333*L2-0.3333*L3

0.3333*L5

-0.3333*L8-0.3333*L9

B*D

1 0

0.3333*L3

0.3333*L6

-0.3333*L8-0.3333*L9

B*D

1 1

-0.3333*L2-0.3333*L3

0.3333*L6

0.3333*L8

B*D

1 2

0.3333*L2

0.3333*L6

0.3333*L9

B*D

2 0

-0.3333*L2-0.3333*L3

-0.3333*L5-0.3333*L6

0.3333*L9

B*D

2 1

0.3333*L2

-0.3333*L5-0.3333*L6

-0.3333*L8-0.3333*L9

B*D

2 2

0.3333*L3

-0.3333*L5-0.3333*L6

0.3333*L8

C*D

0 0

0.3333*L2

0.3333*L5

0.3333*L8

C*D

0 1

-0.3333*L2-0.3333*L3

0.3333*L6

0.3333*L8

C*D

0 2

0.3333*L3

-0.3333*L5-0.3333*L6

0.3333*L8

C*D

1 0

-0.3333*L2-0.3333*L3

-0.3333*L5-0.3333*L6

0.3333*L9

C*D

1 1

0.3333*L3

0.3333*L5

0.3333*L9

C*D

1 2

0.3333*L2

0.3333*L6

0.3333*L9

C*D

2 0

0.3333*L3

0.3333*L6

-0.3333*L8-0.3333*L9

C*D

2 1

0.3333*L2

-0.3333*L5-0.3333*L6

-0.3333*L8-0.3333*L9

C*D

2 2

-0.3333*L2-0.3333*L3

0.3333*L5

-0.3333*L8-0.3333*L9

Type I Estimable Functions - - - - - - - - - - - - - - - - - Coefficients - - - - - - - - - - - - - - - - Effect

D

A*B

A*C

A*D

B*C

B*D

C*D

A*D

1 2

-0.3333*L11-0.3333*L12

0

0

0

0

0

0

A*D

2 0

0.3333*L11

0

0

0

0

0

0

A*D

2 1

0.3333*L12

0

0

0

0

0

0

A*D

2 2

-0.3333*L11-0.3333*L12

0

0

0

0

0

0

574

MAIN EFFECT PLANS

Table 14.2 (Continued ) B*C

0 0

0.3333*L11

0

0

0

0

0

0

B*C

0 1

0.3333*L12

0

0

0

0

0

0

B*C

0 2

-0.3333*L11-0.3333*L12

0

0

0

0

0

0

B*C

1 0

0.3333*L12

0

0

0

0

0

0

B*C

1 1

-0.3333*L11-0.3333*L12

0

0

0

0

0

0

B*C

1 2

0.3333*L11

0

0

0

0

0

0

B*C

2 0

-0.3333*L11-0.3333*L12

0

0

0

0

0

0

B*C

2 1

0.3333*L11

0

0

0

0

0

0

B*C

2 2

0.3333*L12

0

0

0

0

0

0

B*D

0 0

0.3333*L11

0

0

0

0

0

0

B*D

0 1

0.3333*L12

0

0

0

0

0

0

B*D

0 2

-0.3333*L11-0.3333*L12

0

0

0

0

0

0

B*D

1 0

0.3333*L11

0

0

0

0

0

0

B*D

1 1

0.3333*L12

0

0

0

0

0

0

B*D

1 2

-0.3333*L11-0.3333*L12

0

0

0

0

0

0

B*D

2 0

0.3333*L11

0

0

0

0

0

0

B*D

2 1

0.3333*L12

0

0

0

0

0

0

B*D

2 2

-0.3333*L11-0.3333*L12

0

0

0

0

0

0

C*D

0 0

0.3333*L11

0

0

0

0

0

0

C*D

0 1

0.3333*L12

0

0

0

0

0

0

C*D

0 2

-0.3333*L11-0.3333*L12

0

0

0

0

0

0

C*D

1 0

0.3333*L11

0

0

0

0

0

0

C*D

1 1

0.3333*L12

0

0

0

0

0

0

C*D

1 2

-0.3333*L11-0.3333*L12

0

0

0

0

0

0

C*D

2 0

0.3333*L11

0

0

0

0

0

0

C*D

2 1

0.3333*L12

0

0

0

0

0

0

C*D

2 2

-0.3333*L11-0.3333*L12

0

0

0

0

0

0

Dependent Variable: Y Sum of Source

DF

Squares

Mean Square

Model

8

22.22222222

2.77777778

Error

0

0.00000000

.

Corrected Total

8

22.22222222

Source

F Value

Pr > F

.

.

R-Square

Coeff Var

Root MSE

Y Mean

1.000000

.

.

6.555556

DF

Type I SS

Mean Square

F Value

Pr > F

A

2

2.88888889

B

2

2.88888889

1.44444444

.

.

1.44444444

.

.

575


Table 14.2 (Continued ) C

2

1.55555556

0.77777778

.

D

2

14.88888889

7.44444444

.

. .

A*B

0

0.00000000

.

.

.

A*C

0

0.00000000

.

.

.

A*D

0

0.00000000

.

.

.

B*C

0

0.00000000

.

.

.

B*D

0

0.00000000

.

.

.

C*D

0

0.00000000

.

.

.

(see Section 10.2) where the 2 d.f. of BC consist of comparisons among BC0 , BC1 , and BC2 , with BC0 : (βγ )00 + (βγ )12 + (βγ )21 BC1 : (βγ )10 + (βγ )01 + (βγ )22 BC2 : (βγ )20 + (βγ )02 + (βγ )11 and similarly BC02 : (βγ )00 + (βγ )11 + (βγ )22 BC12 : (βγ )10 + (βγ )02 + (βγ )21 BC22 : (βγ )20 + (βγ )01 + (βγ )12 We can write out similar expressions involving the (βδ)j k and (γ δ)k terms. Utilizing these equivalencies in (14.8) leads to the same result as given in (14.7). 14.2.4 Plackett–Burman Designs For 2n factorials a special method of constructing a large number of OMEPs was provided by Plackett and Burman (1946). The method is based on the existence of Hadamard matrices. Definition 14.1 A square matrix H of order N with entries −1, 1 is called a Hadamard matrix of order N if H H = H H = N I N

(14.9)

Condition (14.9) is equivalent to saying that for a Hadamard matrix H the rows are pairwise orthogonal and the columns are pairwise orthogonal. Given a Hadamard matrix H , an equivalent matrix with all elements in the first column equal to +1 can be obtained by multiplying by −1 each element in every row

576

MAIN EFFECT PLANS

of H whose first element is −1. Such a matrix will be called seminormalized. Through a similar operation one can transform a Hadamard matrix H such that all elements in the first row are equal to +1. If the first column and first row contain only +1 elements, then the matrix is said to be normalized. Except for the trivial Hadamard matrices 1 1 H 1 = (1), H 2 = 1 −1 it is known that a necessary condition of the existence of a Hadamard matrix of order N is that N ≡ 0 mod 4 (Paley, 1933). Hadamard matrices for N ≤ 200 have been constructed (see, e.g., Hedayat and Wallis, 1978; Hedayat, Sloane and Stufken, 1999). For N = 2 , for example, H can be constructed as the -fold Kronecker product of H 2 with itself. Hadamard matrices for N = 4t and t = 1, 2, . . ., 8 are given by Hedayat and Wallis (1978) and for t = 1, 2, . . ., 12 by Dey (1985). Consider now a Hadamard matrix H of order N in normalized (or seminormalized) form and delete from it the first column. Denote the resulting matrix and identify the columns with factors F1 , F2 , . . ., FN−1 . We then take by H H∗ =

1 2

+ II H

and consider each row of H ∗ as a treatment combination from the 2N−1 factorial. These N runs constitute an OMEP for N − 1 factors, with N = 4t(t = 1, 2, . . .). The fact that this FFD is an OMEP follows immediately from the properties of a Hadamard matrix in seminormalized (and hence normalized) form, namely for any column (factor) we have the same number of +1 and −1 ele (and hence the same number of +1 and 0 elements in H ∗ , and for ments in H any pair of columns we have the same number of (+1, +1), (+1, −1), (−1, ). FFDs constructed in this manner are called +1), (−1, −1) combinations in H Plackett–Burman designs. Actually, Plackett and Burman (1946) list designs for t = 1, 2, . . ., 25 (except 23 which was given by Baumert, Golomb, and Hall, + II ). In this way the OMEP includes the control, 1962) using H ∗ = 12 (−H that is, the treatment combination (0, 0, . . ., 0). We note here that for N = 2 the Plackett–Burman designs are the same as the Fisher plans for n = N − 1 factors discussed earlier. Also, the Plackett–Burman designs are relevant for the problem of weighing N − 1 light objects in N weighings, a problem first discussed by Yates (1935) for a spring balance and by Hotelling (1944) for a chemical balance. Hotelling (1944) showed that the optimum weighing design was obtained by using Hadamard matrices. Kempthorne (1948) discussed the problem from the viewpoint of factorial experimentation and showed how fractional factorials could be used. For a detailed discussion of weighing designs the reader is referred to Banerjee (1975). In contrast to the Fisher plans the alias structure for the Plackett–Burman designs is rather complicated and difficult to determine (Margolin, 1968; Lin and


577

Draper, 1993). In any case, each main effect is aliased with a long string of 2-factor interactions and no expressions like (14.7) can be written out easily and studied in case the estimate of a main effect turns out to be “large.” Of course, folding-over the Plackett–Burman design leads to a resolution IV design, thereby dealiasing the main effects. In many cases less drastic measures may, however, achieve the goal of dealiasing some main effects, and that may be all that is needed since main effect plans are often used for screening purposes and only some of the main effects may be important; that is, only some of the factors may be active. This leads to investigating the projections of a Plackett–Burman design into smaller, say k, dimensions, where k < n. Here projection simply means to ignore the remaining n − k factors after k factors have been identified, at least in a rough and preliminary sense, as being active. For example, Lin and Draper (1992) have shown that for the 211 Plackett–Burman design in 12 runs the projection to k = 4 dimensions shows that one needs to add only one more run to identify a 24−1 IV design among the 13 runs. Or for k = 5 one needs to add 8 more runs to obtain a 25−1 IV design plus additional runs. Similar results have been obtained for other Plackett–Burman designs (Lin and Draper, 1993). We shall use the 211 Plackett–Burman design in 12 runs, as given in Table 14.3, to make some additional comments about projections and the estimation of main effects and 2-factor interactions: 1. k = 2: Since the Plackett–Burman designs are orthogonal arrays of strength 2 (see Section 14.2.5), it follows that each combination of the levels of two factors occurs equally often. Hence, for the two specified factors we can estimate the main effects and the 2-factor interaction. 2. k = 3: Inspection of the design in Table 14.3 shows that it contains a complete 23 factorial, allowing the estimation of the three main effects and three 2-factor interactions. The 12 runs no longer constitute an orthogonal design and hence the estimators will be correlated. 3. k = 4: It turns out that the available 11 d.f. allow the estimation of the four main effects and their six 2-factor interactions. This is shown in Table 14.3 using the E-option in SAS PROC GLM (keeping in mind that, e.g., EF01 + EJ01 + EL01 = 0). The ANOVA in Table 14.3 also shows that the design is no longer orthogonal as indicated by different numerical values for the type I and type III sums of squares. 14.2.5 Other Methods We mention here briefly two other methods of constructing OMEPs. Addelman and Kempthorne (1961) describe a procedure for constructing an OMEP for an s n factorial in 2s runs, where n = [2(s − 1)/(s − 1) − 1]. This procedure is an extension of the method used for the Fisher plans, using, however, a different correspondence between the factors and factor combinations from the 2 factorial and the n factors for the FFD. This method leads to OMEPs such as N = 18 = 2 · 32 runs for the 37 , N = 54 = 2 · 33 runs for the 325 , or N = 32 = 2 · 42 runs for the 49 factorial.

578

MAIN EFFECT PLANS

Table 14.3 Plackett–Burman Design for 11 Factors in 12 Runs options pageno=1 nodate; data PBdesign; input A B C D E F G H J K L Y; datalines; 1 1 0 1 1 1 0 0 0 1 0 10 0 1 1 0 1 1 1 0 0 0 1 15 1 0 1 1 0 1 1 1 0 0 0 20 0 1 0 1 1 0 1 1 1 0 0 25 0 0 1 0 1 1 0 1 1 1 0 23 0 0 0 1 0 1 1 0 1 1 1 32 1 0 0 0 1 0 1 1 0 1 1 29 1 1 0 0 0 1 0 1 1 0 1 33 1 1 1 0 0 0 1 0 1 1 0 30 0 1 1 1 0 0 0 1 0 1 1 31 1 0 1 1 1 0 0 0 1 0 1 28 0 0 0 0 0 0 0 0 0 0 0 30 ; run; proc print data=PBdesign; title1 'TABLE 14.3'; title2 'PLACKETT-BURMAN DESIGN'; title3 'FOR 11 FACTORS IN 12 RUNS'; run; proc glm data=PBdesign; class A B C D E F G H J K L; model Y=A B C D E F G H J K L; title4 'ANALYSIS OF VARIANCE'; run; proc glm data=PBdesign; class E F J L; model Y=E|F|J|L@2/e; title3 'PROJECTION ON FACTORS E,F,J,L'; title4 'ANALYSIS WITH MAIN EFFECTS AND TWO-FACTOR INTERACTIONS'; run; Obs

A

B

C

D

E

F

G

H

J

K

L

Y

1 2 3 4 5 6 7 8 9 10 11 12

1 0 1 0 0 0 1 1 1 0 1 0

1 1 0 1 0 0 0 1 1 1 0 0

0 1 1 0 1 0 0 0 1 1 1 0

1 0 1 1 0 1 0 0 0 1 1 0

1 1 0 1 1 0 1 0 0 0 1 0

1 1 1 0 1 1 0 1 0 0 0 0

0 1 1 1 0 1 1 0 1 0 0 0

0 0 1 1 1 0 1 1 0 1 0 0

0 0 0 1 1 1 0 1 1 0 1 0

1 0 0 0 1 1 1 0 1 1 0 0

0 1 0 0 0 1 1 1 0 1 1 0

10 15 20 25 23 32 29 33 30 31 28 30

579


Table 14.3 (Continued ) Class Level Information Class

Levels

Values

A

2

0 1

B

2

0 1

C

2

0 1

D

2

0 1

E

2

0 1

F

2

0 1

G

2

0 1

H

2

0 1

J

2

0 1

K

2

0 1

L

2

0 1

Number of observations 12 Dependent Variable: Y

Source

DF

Sum of Squares

Mean Square

Model

11

575.0000000

52.2727273

Error

0

0.0000000

11

575.0000000

Corrected Total

Source A B C D E F G H J K L

F Value .

Pr > F .

.

R-Square

Coeff Var

Root MSE

Y Mean

1.000000

.

.

25.50000

DF

Type I SS

Mean Square

1 1 1 1 1 1 1 1 1 1 1

3.0000000 27.0000000 12.0000000 16.3333333 176.3333333 133.3333333 1.3333333 21.3333333 108.0000000 1.3333333 75.0000000

3.0000000 27.0000000 12.0000000 16.3333333 176.3333333 133.3333333 1.3333333 21.3333333 108.0000000 1.3333333 75.0000000

F Value . . . . . . . . . . .

Pr > F . . . . . . . . . . .

580

MAIN EFFECT PLANS

Table 14.3 (Continued ) Source A B C D E F G H J K L

DF

Type III SS

Mean Square

1 1 1 1 1 1 1 1 1 1 1

3.0000000 27.0000000 12.0000000 16.3333333 176.3333333 133.3333333 1.3333333 21.3333333 108.0000000 1.3333333 75.0000000

3.0000000 27.0000000 12.0000000 16.3333333 176.3333333 133.3333333 1.3333333 21.3333333 108.0000000 1.3333333 75.0000000

F Value . . . . . . . . . . .

Pr > F . . . . . . . . . . .

PROJECTION ON FACTORS E, F, J, L ANALYSIS WITH MAIN EFFECTS AND TWO-FACTOR INTERACTIONS Class Level Information Class

Levels

Values

E

2

0 1

F

2

0 1

J

2

0 1

L

2

0 1

Number of observations

12

General Form of Estimable Functions Effect

Coefficients

Intercept

L1

E E

0 1

L2 L1-L2

F F

0 1

L4 L1-L4

E*F E*F E*F E*F

0 0 1 1

J J

0 1

0 1 0 1

L6 L2-L6 L4-L6 L1-L2-L4+L6 L10 L1-L10

581


Table 14.3 (Continued ) E*J E*J E*J E*J

0 0 1 1

0 1 0 1

L12 L2-L12 L10-L12 L1-L2-L10+L12

F*J F*J F*J F*J

0 0 1 1

0 1 0 1

L16 L4-L16 L10-L16 L1-L4-L10+L16

L L

0 1

E*L E*L E*L E*L

0 0 1 1

0 1 0 1

L22 L2-L22 L20-L22 L1-L2-L20+L22

F*L F*L F*L F*L

0 0 1 1

0 1 0 1

L26 L4-L26 L20-L26 L1-L4-L20+L26

J*L J*L J*L J*L

0 0 1 1

0 1 0 1

L30 L10-L30 L20-L30 L1-L10-L20+L30

L20 L1-L20

Dependent Variable: Y

Source

DF

Sum of Squares

Mean Square

F Value

Pr > F

Model

10

574.5000000

57.4500000

114.90

0.0725

Error

1

0.5000000

0.5000000

11

575.0000000

Corrected Total

Source E F

R-Square

Coeff Var

Root MSE

Y Mean

0.999130

2.772968

0.707107

25.50000

DF

Type I SS

Mean Square

F Value

Pr > F

1 1

176.3333333 133.3333333

176.3333333 133.3333333

352.67 266.67

0.0339 0.0389

582

MAIN EFFECT PLANS

Table 14.3 (Continued ) E*F J E*J F*J L E*L F*L J*L Source E F E*F J E*J F*J L E*L F*L J*L

1 1 1 1 1 1 1 1

65.3333333 66.6666667 2.6666667 112.6666667 10.8000000 5.4857143 0.1758242 1.0384615

65.3333333 66.6666667 2.6666667 112.6666667 10.8000000 5.4857143 0.1758242 1.0384615

130.67 133.33 5.33 225.33 21.60 10.97 0.35 2.08

0.0556 0.0550 0.2601 0.0423 0.1349 0.1867 0.6592 0.3862

DF

Type III SS

Mean Square

F Value

Pr > F

1 1 1 1 1 1 1 1 1 1

61.03846154 61.03846154 9.34615385 61.03846154 0.11538462 72.11538462 9.34615385 5.65384615 0.11538462 1.03846154

61.03846154 61.03846154 9.34615385 61.03846154 0.11538462 72.11538462 9.34615385 5.65384615 0.11538462 1.03846154

122.08 122.08 18.69 122.08 0.23 144.23 18.69 11.31 0.23 2.08

0.0575 0.0575 0.1447 0.0575 0.7149 0.0529 0.1447 0.1840 0.7149 0.3862

Another method is derived from the fact that each OMEP is an orthogonal array of strength 2 (Rao, 1946a). More precisely, if n denotes the number of factors, N the number of runs, s the number of levels, then an OMEP can be denoted by OA[N , n, s, 2]. Bose and Bush (1952) show that an OA[λs 2 , λs, s, 2] can be constructed if λ and s are both powers of the same prime p. Suppose λ = p u , s = p ν . Then some examples of OMEPs obtained with this method are given below: Factorial

Number of Runs (N )

24 28 39 327

8 16 27 81

Obviously, these will not be saturated OMEPs. 14.3 ORTHOGONAL RESOLUTION III DESIGNS FOR ASYMMETRICAL FACTORIALS In many practical situations it is the case that we shall use factors with different numbers of levels. This is dictated by the nature of the experiment and the

583

ORTHOGONAL RESOLUTION III DESIGNS FOR ASYMMETRICAL FACTORIALS

type of factors included in the experiment as well as the suspected nature of their response at different levels. And again, if we want to investigate a large number of factors, concentrating on estimating main effects may be a prudent thing to do. 14.3.1 Kronecker Product Design n

An easy method to construct an OMEP for an s1n1 × s2n2 × · · · × sq q asymmetrical factorial is to consider each component symmetrical factorial separately, construct an OMEP for it, and then take the Kronecker product of these designs. For 4−2 example, if we have a 27 × 34 we can construct a 27−4 III and a 3III and the Kronecker product of these two designs will yield a FFD, which is an OMEP, with N = 8 · 9 = 72 runs. Obviously, this design contains far too many treatment combinations to account for the 7 + 4 · 2 = 15 d.f. for the main effects. It is for this reason that other methods of constructing OMEPs for asymmetrical factorials have been developed. 14.3.2 Orthogonality Condition A method that yields a very large number of OMEPs for many different situations was developed by Addelman and Kempthorne (1961) (see also Addelman, 1962a). These designs are based on OMEPs for symmetrical factorials with subsequent collapsing and replacement of levels for certain factors. This procedure will be described more precisely in Section 14.3.3. For now we mention only that as a consequence of this procedure the levels of one factor no longer occur together with the same frequency with each of the levels of any other factor, as was the case for the OMEPs for symmetrical factorials. They do, however, occur together with proportional frequencies and that, as we shall show, is sufficient to guarantee orthogonal estimates of main effects. To reduce the amount of writing let us consider the case of two factors, A and B say, where A has a levels and B has b levels. Suppose that in a given FFD the ith level of A occurs together the j th level of B for Nij times. with Let j Nij = Ni. , i Nij = N.j , i Ni. = j N.j = N , the total number of runs in the FFD. If the level combinations occur with proportional frequencies, we have Nij = Ni. N.j /N for every i and j (i = 0, 1, . . ., a − 1; j = 0 1, . . ., b − 1). We then state the following theorem. Theorem 14.1 A sufficient condition for a main effect plan for two factors A and B with a and b levels, respectively, to be an OMEP is that each level combination for the two factors occurs with proportional frequency. Proof Consider model (13.28) and rewrite it as   µ   E(y) = (I, X 11 , X12 ) A B

584

MAIN EFFECT PLANS

where A = (A0 , A1 , . . . , Aa−1 ) and B = (B0 , B1 , . . . , Bb−1 ) represent the main effect components for the two factors A and B, respectively. We then rewrite the N E (13.29) as 

I I

I X11

X12 I

X 12 X 11

    µ Iy      X 11 X12  A = X11 y  I X12

 X11 I X 11 X 11

X 12 X12

(14.10)

X12 y

B

Denote the coefficient matrix in (14.10) by K. Then we have 

N

 N  0.   N1.   .  .  .  K= Na−1,.   N.0   N  .1   ..  .

N0.

N1.

...

Na−1,.

N.0

N.1

...

N0.

0

...

0

N00

N01

...

0

N1.

...

0

N10

N11

...

N.,b−1



N0,b−1  



N1,b−1  

    . . . Na−1,b−1    ... 0    ... 0    ..  . .. .

0

0

...

Na−1,.

N00

N10

...

Na−1,0

N.0

0

N01

N11

...

Na−1,1

0

N.1

0

0

Na−1,0 Na−1,1

N.,b−1 N0,b−1 N1,b−1 . . . Na−1,b−1

...

N.,b−1

To obtain a solution to (14.10) we need to find a generalized inverse for K. If Nij = Ni. N.j /N for each i and j , then such a g inverse is given by 

1 N 1 − N 1 − N .. .

              1  K− =  − N   1 −  N   1 −  N   .  ..    1 −

N

 0

0

...

0

0

...

1 N0. 1 N1.

0 ..

. 1 Na−1,. 1 N.0 1 N.1

0

..

. 1 N.,b−1

                (14.11)              


585

This follows easily by verifying that KK − K = K using the proportionality con− dition. The reader may recognize that the solution and hence K can be obtained

Ai = 0 = j N.j Bj . Consider now the conby imposing the conditions i Ni. dj Bj (with ci = 0, dj = 0) belonging to the main trasts ci Ai and effects A and B, respectively. It follows then immediately that   0

i ,

j = (0, c , 0)K −  0  = 0 (14.12) cov ci A dj B d where c = (c0 , c1 , . . . , ca−1 ), d = (d0 , d1 , . . . , db−1 ), that is, estimates of contrasts for the two main effects are uncorrelated. Hence we have an OMEP. Obviously the proof can be extended in a straightforward manner to any number of factors with the proportionality condition holding for any pair of factors. We mention also that Addelman and Kempthorne (1961) proved that proportional frequency is actually a necessary and sufficient condition for a main effect plan to be an OMEP (see also Dey, 1985). The methods of constructing main effect plans as described in the following section make use of the proportionality condition of Theorem 14.1. Before proceeding, however, we need to clarify an argument put forth by Srivastava and Ghosh (1996). They argue that main effect plans with proportional (and not equal) frequencies are not OMEPs. Clearly, Eq. (14.12) shows that the estimators of contrasts of main effects belonging to two different factors are uncorrelated. The concern of Srivastava and Ghosh (1996) has to do with estimators of contrasts belonging to the same factor, for example,

i

i ci A di A (14.13)

i

i

with i ci = 0 = i di and i ci di = 0. Using K − from (14.11), it follows immediately that the contrasts in (14.13) are correlated (see also Section I.7.3). In order to produce uncorrelated contrasts Addelman (1962a) considers contrasts of the form Ni. ci∗ Ai Ni. di∗ Ai (14.14)

i

i

with i Ni. ci∗ = 0 = i Ni. di∗ and i Ni. ci∗ di∗ = 0. We emphasize that contrasts of the form (14.13) or (14.14) are considered only for purposes of posthoc analyses (see Chapter I.7). For qualitative factors such contrasts are prespecified, for quantitative factors they represent various forms of trends, such as linear, quadratic, and so forth. In either case the contrasts must be chosen such that they are meaningful in the context of the experiment, whether they are orthogonal or not. And that does not affect the properties of the design. Whether a main effect plan is an OMEP is determined only by whether the condition of Theorem 14.1 is satisfied.

586

MAIN EFFECT PLANS

14.3.3 Addelman–Kempthorne Methods We now describe three methods for constructing OMEPs for asymmetrical factorials as developed by Addelman and Kempthorne (1961). Method 1 n We consider the s1n1 × s2n2 × · · · × sq q factorial and assume that s1 > s2 > · · · > sq . Collapsing Factors We can construct an OMEP with N = s1 runs, where s1 is a prime or prime power and q

ni ≤

i=1

s1 − 1 s1 − 1

(14.15)

in the following way. We first construct an OMEP for the s1n factorial in s1 trials where is chosen so that (14.15) is satisfied and where n = (s1 − 1)/(s1 − 1) using the method described in Sections 14.2 and 11.7. We write the plan as an N × n array, D say, such that the columns denote the factors and each row represents a treatment combination (run) for the n factors with s1 levels each. We shall refer to D also as the auxiliary OMEP. In the context of the s1n1 × q s2n2 × · · · × sqn factorial the first n1 columns in D correspond to the first set of factors, each having s1 levels. For the following factors we use what Addelman and Kempthorne (1961) refer to as collapsing factors with s1 levels to factors occurring at si levels (i = 2, 3, . . ., q) by setting up a many–one correspondence between the set of s1 levels and the set of si levels. This means that different levels of the s1 -level factor are mapped into the same level of the si -level factor. As an example, suppose we have s1 = 5, s2 = 3. We then might have the following mapping: Five-Level Factor

Three-Level Factor

0

−→

0

1

−→

1

2

−→

2

3

−→

0

4

−→

1

Proceeding in this way we replace the next n2 factors by s2 -level factors, the following n3 factors by s3 -level factors, and so on, until all factors have been q accommodated. Since i=1 ni ≤ n it may be necessary to delete some factors from the basic OMEP if the inequality holds. Since the collapsing procedure leads automatically to proportional frequencies (for any collapsing scheme) the final design is an OMEP. We illustrate this procedure with the following example.


587

Example 14.2 Consider the 53 × 3 × 22 factorial. Since for = 2, (52 − 1)/(5 − 1) = 6 we can construct an OMEP with 25 runs from an OMEP for the 56 factorial in 25 runs, using the method of Section 11.7.1. The construction of this plan is illustrated in Table 14.4a. Using the collapsing schemes 0

−→

0

0

−→

0

1

−→

1

1

−→

1

2

−→

2

2

−→

0

3

−→

0

3

−→

1

4

−→

1

4

−→

0

and

we then obtain the final design given in Table 14.4b. It can easily be verified that the proportionality conditions are met. For example, for factors F4 and F5 we have N00 = 6, N0. = 10, N.0 = 15, and with N = 25 we have N00 = 10 · 15/25 = 6. Expanding Factors Rather than collapsing the s1 -level factor to an si -level factor we may, in certain situations, expand the s1 -level factor into a set of si -level factors if s1 = sim for some i. We know that there exists an OMEP for ν = (sim − 1)/(si − 1) factors ν−(ν−m) n−(n−) at si levels with sim runs, that is, an siIII . We then construct the s1III for a suitable n as discussed before and determined from n ≤ (s1 − 1)/(s1 − 1). Next, we replace each of the s1 levels for one factor by a treatment combination ν−(ν−m) from the siIII . We thus replace an s1 -level factor by ν = (sim − 1)/(si − 1) si -level factors. This method will be most useful for s1 = 4 = 22 and perhaps for s1 = 8 = 23 , s1 = 9 = 32 . We now illustrate this method in the following example. Example 14.3 Consider the 43 × 3 × 23 factorial. Instead of using the collapsing factors method with n1 = 3, n2 = 1, n3 = 3 which would lead to an OMEP with 64 runs, we use the fact that a 4-level factor can be replaced by ν = 3 two-level factors. We thus consider an OMEP for the 45 factorial that can be obtained with 16 runs and then use replacing and collapsing as given in Table 14.5. For details of the construction of the auxiliary OMEP using GF(22 ) arithmetic, we refer to Sections 11.7 and 11.13.1 (Example 11.5). In the final plan (Table 14.5b), the elements in GF(22 ) are replaced by 0, 1, 2, and 3, respectively. We use the following collapsing scheme for F4 : Four-Level Factor (F4 ) 0 1 2 3

Three-Level Factor (F4 ) −→ −→ −→ −→

0 1 2 0

588

MAIN EFFECT PLANS

Table 14.4 Orthogonal Main Effect Plan for 53 × 3 × 22 Factorial α x

F1

F2

F3

F4

F5

F6

(1 0)

(0 1)

(1 1)

(1 2)

(1 3)

(1 4)

0 2 4 1 3 1 3 0 2 4 2 4 1 3 0 3 0 2 4 1 4 1 3 0 2

0 3 1 4 2 1 4 2 0 3 2 0 3 1 4 3 1 4 2 0 4 2 0 3 1

0 4 3 2 1 1 0 4 3 2 2 1 0 4 3 3 2 1 0 4 4 3 2 1 0

a. Auxiliary Plan (56 in 25 Runs) 0 0 0 0 0 1 1 1 1 1 2 2 2 2 2 3 3 3 3 3 4 4 4 4 4

0 1 2 3 4 0 1 2 3 4 0 1 2 3 4 0 1 2 3 4 0 1 2 3 4

0 0 0 0 0 1 1 1 1 1 2 2 2 2 2 3 3 3 3 3 4 4 4 4 4

0 1 2 3 4 0 1 2 3 4 0 1 2 3 4 0 1 2 3 4 0 1 2 3 4

0 1 2 3 4 1 2 3 4 0 2 3 4 0 1 3 4 0 1 2 4 0 1 2 3 b. Final Plan

F1

F2

F3

F4

F5

F6

0 0 0 0 0 1 1 1 1 1 2

0 1 2 3 4 0 1 2 3 4 0

0 1 2 3 4 1 2 3 4 0 2

0 2 1 1 0 1 0 0 2 1 2

0 1 1 0 0 1 0 0 0 1 0

0 0 1 0 1 1 0 0 1 0 0

589


Table 14.4 (Continued ) F1

F2

F3

F4

F5

F6

2 2 2 2 3 3 3 3 3 4 4 4 4 4

1 4 3 4 0 1 2 3 4 0 1 2 3 4

3 4 0 1 3 4 0 1 2 4 0 1 2 3

1 1 0 0 0 0 2 1 1 1 1 0 0 2

0 1 1 0 1 1 0 0 0 0 0 0 1 1

1 0 0 1 1 0 1 0 0 0 1 0 1 0

Table 14.5 Orthogonal Main Effect Plan for 43 × 3 × 23 Factorial α x

F1

F2

F3

F4

F5

(u1 , u0 )

(u0 , u1 )

(u1 , u1 )

(u1 , u2 )

(u1 , u3 )

u0 u2 u3 u1 u1 u3 u2 u0 u2 u0 u1 u3 u3 u1 u0 u2


a. Auxiliary OMEP u0 , u0 , u0 , u0 , u1 , u1 , u1 , u1 , u2 , u2 , u2 , u2 , u3 , u3 , u3 , u3 ,





590

MAIN EFFECT PLANS


F1

F2

F3

b. Final Plan F4

0 0 0 0 1 1 1 1 2 2 2 2 3 3 3 3

0 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3

0 1 2 3 1 0 3 2 2 3 0 1 3 2 1 0

0 2 0 1 1 0 2 0 2 0 1 0 0 1 0 2

F5,1

F5,2

F5,3

0 1 0 1 0 1 0 1 1 0 1 0 1 0 1 0

0 1 1 0 1 0 0 1 0 1 1 0 1 0 0 1

0 0 1 1 1 1 0 0 1 1 0 0 0 0 1 1

and the following replacement (expansion): F5

F5,1

F5,2

F5,3

0

−→

0

0

0

1

−→

0

1

1

2

−→

1

0

1

3

−→

1

1

0

Method 2 n For the s × s1n1 × s2n2 × · · · × sq q factorial we can construct an OMEP with 2s1 runs, where s1 is a prime or prime power and s1 > s2 > · · · > s1 and 1+

s1 < s ≤ 2s1 q i=1

ni ≤

s1 − 1 s1 − 1

(14.16)

We construct first the OMEP for the s1ν factorial where ν = (s1 − 1)/(s1 − 1) and is chosen such that (14.16) is satisfied. We then duplicate the plan except

591


that the levels 0, 1, 2, . . ., s1 − 1 for the first factor, F1 , are replaced by s1 , s1 + 1, s1 + 2, . . ., 2s1 − 1, respectively. This will be an OMEP with 2s1 runs for the 2s1 × s1ν factorial. If s < 2s1 , we obtain the final plan by collapsing F1 and then proceed as in method 1a. We illustrate this with the following example. Example 14.4 We consider the 5 × 32 × 2 factorial. We have 1 + 2 + 1 = 4 ≤ (3 − 1)/(3 − 1). Hence = 2, and we construct the 34−2 III , duplicate it (see Table 14.6a) and collapse factors F1 and F4 as follows: Six-Level Factor (F1 )

Five-Level Factor (F1 ) −→ −→ −→ −→ −→ −→

0 1 2 3 4 5

0 1 2 3 4 0

Three-Level Factor (F4 )

Two-Level Factor (F4 ) −→ −→ −→

0 1 2

0 1 0

The final design is given in Table 14.6b.

Method 3 In some cases the methods described above lead to an unnecessarily large number of runs since the auxiliary plan is based on an OMEP involving factors with the largest number of levels. The method to be described is, by contrast, based on an appropriate OMEP for factors having the smallest number of levels. The basic idea is, as a converse to method 1b, to replace several factors with the smallest number of levels by a factor with a larger number of levels. More precisely, n we can construct an OMEP for the s1n1 × s2n2 × · · · × sq q factorial with s1 runs, where s1 is a prime or prime power, s1 < s2 , . . . , < sq and q i=1

λi ni ≤

s1 − 1 s1 − 1

(14.17)

with 1 = λ1 < λ2 ≤ λ3 ≤ · · · ≤ λq and λi (i = 1, 2, . . ., q) an integer. The basic idea of this procedure is based on the following theorem. n−(n−) Theorem 14.2 In the sIII with n = (s − 1)/(s − 1) and s a prime or prime power, a factor with t levels, s m−1 < t ≤ s m , can be introduced as a replacement for a suitably chosen set of ν = (s m − 1)/(s − 1) factors that will preserve the orthogonality of the main effect estimates.

592

MAIN EFFECT PLANS

Table 14.6 Orthogonal Main Effect Plan for 5 × 32 × 2 Factorial a. Auxiliary Plan

b. Final Plan

F1

F2

F3

F4

F1

F2

F3

F4

0 0 0 1 1 1 2 2 2 3 3 3 4 4 4 5 5 5

0 1 2 0 1 2 0 1 2 0 1 2 0 1 2 0 1 2

0 1 2 1 2 0 2 0 1 0 1 2 1 2 0 2 0 1

0 2 1 1 0 2 2 1 0 0 2 1 1 0 2 2 1 0

0 0 0 1 1 1 2 2 2 3 3 3 4 4 4 0 0 0

0 1 2 0 1 2 0 1 2 0 1 2 0 1 2 0 1 2

0 1 2 1 2 0 2 0 1 0 1 2 1 2 0 2 0 1

0 0 1 1 0 0 0 1 0 0 0 1 1 0 0 0 1 0

ν−(ν−m) Proof Let t = s m . There exists an sIII with ν = (s m − 1)/(s − 1). Hence a factor with t = s m levels can replace ν factors with s levels each. If s m−1 < t < s m , then a factor with s m levels can be collapsed. We illustrate this method with the following example.

Example 14.5 We consider the 24 × 4 factorial, that is, s1 = 2, s2 = 4, n1 = 4, n2 = 1. Now t = 4 = 22 , that is, m = 2 and ν = 3 and a 4-level factor can replace three 2-level factors in the 2n−(n−) where n and are determined as III follows so that (14.17) is satisfied: With n1 = 4, n2 = 1, and ν = 3 we have for (14.17) with λ2 = ν = 3 4 + 3 · 1 = n ≤ (2 − 1)/(2 − 1) which yields = 3. Hence we can obtain the OMEP with N = 23 = 8 runs from the 27−4 III . The final plan is given in Table 14.7 using the replacement according to the 23−1 III given below: Two-Level Factors in 23−1 III 0 0 1 1

0 1 0 1

0 1 1 0

Four-Level Factor −→ −→ −→ −→

0 1 2 3

593


Table 14.7 Orthogonal Main Effect Plan for 24 × 4 Factorial α x

F1

F2

F3

F4

F5

F6

F7

(1 0 0)

(0 1 0)

(1 1 0)

(0 0 1)

( 1 0 1)

(0 1 1)

(1 1 1)

0 1 0 1 1 0 1 0

0 1 1 0 0 1 1 0

0 1 1 0 1 0 0 1

a. Auxiliary Plan 0 0 0 0 1 1 1 1

0 0 1 1 0 0 1 1

0 1 0 1 0 1 0 1

0 0 0 0 1 1 1 1

0 0 1 1 0 0 1 1

0 0 1 1 1 1 0 0

0 1 0 1 0 1 0 1

b. Final Plan F1

F4

F5

F6

F7

0 0 1 1 2 2 3 3

0 1 0 1 0 1 0 1

0 1 0 1 1 0 1 0

0 1 1 0 0 1 1 0

0 1 1 0 1 0 0 1

It follows then that in the auxiliary plan (Table 14.5a) factors F1 , F2 , F3 with two levels are replaced by factor F1 with four levels. This procedure of constructing an OMEP is most useful if s < t ≤ s 2 . It follows then from Theorem 14.2 that the maximum number of t-level factors n−(n−) that can be introduced into the sIII with n = (s − 1)/(s − 1) is equal to (s − 1)/(s − 1) if is even, and equal to the largest integer less than or equal to (s − 1)/(s − 1) − 1 if is odd. The three methods described above lead to a very large number of OMEPs for asymmetrical factorials. An extensive catalog of such plans is provided by Addelman and Kempthorne (1961), Addelman (1962a, 1962b), and Dey (1985). Dey also lists some additional OMEPs not derivable by the Addelman–Kempthorne methods but instead derivable from Hadamard matrices or orthogonal arrays. For details the reader is referred to Dey (1985). For the 2n × 3m Addelman–Kempthorne designs Margolin (1968) has investigated the aliasing of main effects with 2-factor interactions. This is important for the reasons discussed in Section 14.2 for OMEPs for symmetrical factorials. We shall not discuss the rather complicated results here but refer to the method illustrated in Table 14.2.

594

MAIN EFFECT PLANS

14.4 NONORTHOGONAL RESOLUTION III DESIGNS Orthogonal main effect plans are useful because they are easy to analyze and the results are easy to interpret. Sometimes there may, however, be practical reasons why an OMEP may not be the most suitable design. In that case nonorthogonal MEPs may be more appropriate. We shall give some examples. As we have mentioned earlier, a saturated OMEP does not provide any degrees of freedom for the estimation of the error variance. If no information about the error variance is available, additional observations must be obtained. To replicate the OMEP may lead to too large an experiment. Hence only some runs may be replicated leading to nonproportional frequencies and as a consequence to correlated estimates of the main effects. In contrast to the just described situation an OMEP may not be a saturated plan and may therefore have too many runs. It is always possible to obtain a saturated MEP. One such plan is the one-at-a-time plan, where if the runs are considered in sequence, only the level of one factor is changed from one run to the next. As an example consider the following MEP for a 28 in nine runs (apart from randomization): Run

F1

F2

F3

F4

F5

F6

F7

F8

1 2 3 4 5 6 7 8 9

0 1 1 1 1 1 1 1 1

0 0 1 1 1 1 1 1 1

0 0 0 1 1 1 1 1 1

0 0 0 0 1 1 1 1 1

0 0 0 0 0 1 1 1 1

0 0 0 0 0 0 1 1 1

0 0 0 0 0 0 0 1 1

0 0 0 0 0 0 0 0 1

An alternative to the plan above is one where each run, other than the control, has only one factor at the high level and all other factors at the low level (apart from randomization): Run

F1

F2

F3

F4

F5

F6

F7

F8

1 2 3 4 5 6 7 8 9

0 1 0 0 0 0 0 0 0

0 0 1 0 0 0 0 0 0

0 0 0 1 0 0 0 0 0

0 0 0 0 1 0 0 0 0

0 0 0 0 0 1 0 0 0

0 0 0 0 0 0 1 0 0

0 0 0 0 0 0 0 1 0

0 0 0 0 0 0 0 0 1

NONORTHOGONAL RESOLUTION III DESIGNS

595

Obviously, this method can be adapted to both symmetrical and asymmetrical factorials with factors at any number of levels. It should be clear to the reader that if a saturated OMEP exists for a given factorial it will be more efficient than the comparable nonorthogonal MEP. But often statistical properties may have to be balanced against practical considerations. In listing the designs above we have added for emphasis the cautionary note “apart from randomization.” There may be settings, especially in industrial applications, where it would be most economical and feasible to execute the experiment using the run order given above. This may have serious consequences for the analysis and interpretation, since not only is there lack of randomization (see I.9.9) but this would also invoke a split-plot type of experiment if the levels of factors are not reset for each run (see I.13). For a discussion of such problems see Daniel (1973) and Webb (1968).

C H A P T E R 15

Supersaturated Designs

15.1 INTRODUCTION AND RATIONALE We have seen in the preceding chapters that very often practical considerations lead the experimenter to using fractional factorial designs. The aim, of course, is to reduce the size of the experiment without sacrificing important information. Even then, the suggested design may be quite large. This may be true even for the most highly fractionated design, the saturated main effect plan, if the number of factors under investigation is large. And it is not unusual that in many physical, chemical, or industrial experiments the number of factors that one would like to study is quite large, but consideration of cost calls for a small number of runs, that is, treatment combinations. Moreover, it is suspected that among the large number of factors only few are important or active. Such situations have led to the development of what are called supersaturated designs (for a definition see Section 15.3). The idea is to somehow select N treatment combinations from the totality of all possible treatment combinations, say from the s1 × s2 × · · · × sn combinations of a factorial with factors A1 , A2 , . . . , An where factor Ai has pi levels (i = 1, 2, . . . , n) and n > N − 1. Obviously, orthogonality as with fractional factorials can no longer be achieved. Rather, orthogonality is being replaced by some measure of near orthogonality. Also the analysis of data from such an experiment will have to be approached differently from the usual analysis of factorial experiments. Simultaneous estimation of effects will no longer be possible. Instead, based on the assumption of scarcity of active factors and absence of interactions, stepwise forward regression provides a plausible approach. 15.2 RANDOM BALANCE DESIGNS One way to select the N treatment combinations is by using a random sampling process. This idea has led to the notion of random balance designs (Satterthwaite, 1959). Design and Analysis of Experiments. Volume 2: Advanced Experimental Design By Klaus Hinkelmann and Oscar Kempthorne ISBN 0-471-55177-5 Copyright  2005 John Wiley & Sons, Inc.

596

DEFINITION AND PROPERTIES OF SUPERSATURATED DESIGNS

597

Here the notion of random balance is understood in the following sense (Raghavarao, 1986): Factor Aj is randomly balanced with respect to factor Ai if the random sampling process used to select the pj levels of factor Aj is the same for each level of factor Ai . If every factor has random balance with respect to every other factor, then Satterthwaite (1959, p. 113) refers to this as a pure random balance design. Satterthwaite (1959, p. 114) states that “pure random balance designs are almost always inefficient in some sense” (see Section 15.4.1), but at the same time he argues that “there are important and large classes of problems for which these inefficiencies are often unimportant or even trivial.” These views were largely contradicted by Youden et al. (1959), who, among other things, called for a systematic, rather than random, construction of designs for the earlier stated purpose.

15.3 DEFINITION AND PROPERTIES OF SUPERSATURATED DESIGNS Generalizing some results of Yamada and Lin (1999), we consider an experiment with n factors A1 , A2 , . . . , An , each with p levels denoted by 1, 2, . . . , p. Let N be the number of runs (treatment combinations) with (i) (p − 1)n > N − 1 and (ii) N = p · q and q a positive integer. Further, let C(N ) be the set of N -dimensional vectors each containing each of the p levels exactly q times. A p-level supersaturated design D can be described as a selection of n vectors in C(N ), say c1 , c2 , . . . , cn , by some specified rule. The design D can then be represented by the N × n array [c1 , c2 , . . . , cn ]

(15.1)

Obviously, the rule by which the selection leading to (15.1) is made is of major importance. As mentioned earlier, orthogonality of the vectors in (15.1) cannot be achieved, but we would like to stay as “close” as possible to this property, which is exhibited in the Fisher and Plackett–Burman main effect plans (see Sections 14.2.1 and 14.2.4). Following Yamada and Lin (1999) we shall adopt the following χ 2 statistic as a measure of deviation from orthogonality and choose the design D, which achieves the smallest amount of deviation. Such a design we declare to exhibit near-optimality. Let nab (ci , cj ) denote the number of rows in the N × 2 array [ci , cj ] whose values are [a, b], where ci , cj , (i < j ) are two columns from (15.1). Obviously, p a,b=1

nab (ci , cj ) = N

598

SUPERSATURATED DESIGNS

for any pair i, j (1 ≤ i < j ≤ n). The χ 2 statistic mentioned above is then defined as χij2

p [nab (ci , cj ) − N/p 2 ]2 = N/p 2

(15.2)

a,b=1

For orthogonality (or complete independence) of ci and cj we have nab (ci , cj )= N/p 2 for each pair [a, b] (1 ≤ a, b ≤ p), whereas for complete confounding of factors Ai , Aj (or complete dependence) we have, without loss of generality, n (ci , cj ) = ab

N/p 0

for a = b otherwise

It is then easy to see that the χ 2 value of (15.2) lies between 0 and (p − 1)N . For the overall evaluation of deviation from orthogonality for the design D, we then use either n

av χ 2 =

2 χi,j

i,j =1 i<j

(15.3)

n(n − 1)/2

or max χ 2 = max χij2

1≤i<j ≤n

(15.4)

with χij2 given in (15.2). Using either (15.3) or (15.4) the design D with the smallest value will be chosen among competing arrangements. For an extension of the χ 2 criterion to mixed-level supersaturated designs see Yamada and Matsui (2002).

15.4 CONSTRUCTION OF TWO-LEVEL SUPERSATURATED DESIGNS 15.4.1 Computer Search Designs Most of the emphasis of devising rules for obtaining a suitable array (15.1) has been concentrated on 2-level supersaturated designs, that is, p = 2, denoting the levels by −1 and +1. Booth and Cox (1962) provided the first set of such designs for the following pairs of (n, N ) values: (16, 12), (20, 12), (24, 12), (24, 18), (30, 18), (36, 18), and (30, 24). These designs were obtained through computer search.

CONSTRUCTION OF TWO-LEVEL SUPERSATURATED DESIGNS

599

To compare their designs with random balance designs, Booth and Cox (1962) consider sij = ci cj (i < j ) as a measure of nonorthogonality of factors Ai and Aj . As an overall measure of the non-orthogonality for the design under consideration, they then define sij2 E(s 2 ) =

i<j

n(n − 1)/2

(15.5)

Yamada and Lin (1999) show that (15.5) is equivalent to (15.3) for p = 2. For a random balance design (15.5) is given by ERBD (s 2 ) =

N2 N −1

(15.6)

(Booth and Cox, 1962). It turns out that for all designs constructed by Booth and Cox (1962) E(s 2 ) < ERBD (s 2 ) indicating a smaller degree of nonorthogonality, that is, less confounding among the factors, of their designs, a result that holds true in general for other supersaturated designs as well. 15.4.2 Hadamard-Type Designs A more systematic way to construct a special class of 2-level supersaturated designs was provided by Lin (1993). Just as the Plackett–Burman designs (see Section 14.2.4) are based on Hadamard matrices of order M, so are the proposed supersaturated designs. Their construction can be described in general as follows: Consider a Hadamard matrix of order M (which is not constructed via foldover) in seminormalized form. Delete the first column of +1’s and label the remaining columns 1, 2, . . . , M − 1. Choose one of these columns as the branching column, that is, the column that divides the Hadamard matrix into two half-fractions according to its +1’s and −1’s. Delete the branching column and choose a sign, say +. Then the M/2 rows that corresponding to +1 in the branching column form a supersaturated design with n = M − 2 factors and N = M/2 runs, where each factor is, of course, represented M/4 times at the high and low level. Letting H be an M × M Hadamard matrix, then the procedure described above can be formally represented as I I Hh H = (15.7) I −I ∗

600


Table 15.1 Supersaturated Design for n = 10 Factors in N = 6 Runs Factor Row 1 2 3 4 5 6 7 8 9 10 11 12

+ + + + + + + + + + + +

+ + + + + + − − − − − −

1

2

3

4

5

6

7

8

9

10

Run

+ − + + − − + + − − + −

− + + − − + + + − + − −

+ + − − − + − + + − + −

+ + − − + − + − − + + −

+ − − + − + + − + + − −

− − + − + + + − + − + −

− − − + + + − + − + + −

− + + + − − − − + + + −

+ − + − + − − + + + − −

− + − + + − + + + − − −

1 2 3 4 5 6

(Bulutoglu and Cheng, 2003), where I represents a vector of N/2 unity elements and H h represents the supersaturated design, that is, the columns of H h represent the columns [c1 , c2 , . . . , cn ] of (15.1) with n = M − 2. Example 15.1 We shall illustrate this method with the design for M = 12 given by Lin (1993). Table 15.1 contains the Hadamard matrix in the form indicated by (15.7), where + represents +1 and − represents −1. It is easy to see that for this from (15.5) design all | sij | = 2 and hence E(s 2 ) = 4, compared to ERBD (s 2 ) = 7.2 from (15.6). For n < M − 2 we can delete certain columns from the complete design. In practice it does not matter which column we choose as the branching column and which columns we delete subsequently (Lin, 1993). Another Hadamard-type supersaturated design was proposed by Wu (1993). Starting with a seminormalized Hadamard matrix of order M, it makes use of all its columns except the first and augments it by interaction columns to generate a supersaturated design with N = M runs for n factors, where M ≤ n ≤ 2M − 3. The procedure can be described as follows. Let the generator Hadamard matrix H be given by H = (I, x 1 , x 2 , . . . , x M−1 ). Obtain M − 2 interaction columns x i0 j = x i0 ◦ x j as the elementwise product of x i0 and x j (j = i0 , i0 fixed). For cyclic Hadamard matrices we can choose i0 = 1; for noncyclic Hadamard matrices we choose i0 to minimize max | x i x i0 j | over i and j . Thus, (15.1) is given by [x 1 , x 2 , . . . , x M−1 , x i0 j (j = 1, 2, . . . , M − 1; j = i0 )]

CONSTRUCTION OF TWO-LEVEL SUPERSATURATED DESIGNS

601

Among all inner products of the columns only x i x i0 j with i, i0 , j all different may be nonzero. Hence the choice of i0 ensures that the confounding among factors is kept to a minimum. It can be shown (Wu, 1993) that for the above designs E(s 2 ) =

M 2 (n − M + 1) n(n − 1)/2

(15.8)

If n < 2M − 3, we simply choose n − M + 1 columns from the M − 2 interaction columns, keeping E(s 2 ) as small as possible. A third method to construct Hadamard-type supersaturated designs was proposed by Tang and Wu (1997). The method essentially consists of adjoining k Hadamard matrices (after deleting their first column of ones) of order M. The resulting array yields a supersaturated design for n = k(M − 1) factors in M runs. For n = (k − 1)(M − 1) + j factors (with 1 ≤ j ≤ M − 2) we delete M − 1 − j from the full design, that is, the design for k(M − 1) factors. We shall conclude this section by making some remarks about the possible optimality with regard to the E(s 2 ) criterion. Nguyen (1996) and Tang and Wu (1997) have shown that a lower bound for E(s 2 ) is given by E(s 2 ) ≥

M 2 (n − M + 1) (n − 1)(M − 1)

(15.9)

Nguyen (1996) has shown that the half-fraction Hadamard designs of Lin (1993) achieve the lower bound (15.9). This is true also for the full designs given by Tang and Wu (1997). Comparing (15.8) with (15.9) shows that for n = 2M − 3 the designs of Wu (1993) have E(s 2 ) values only slightly larger than the lower bound. A similar statement can be made about the designs of Tang and Wu (1997) when n = (k − 1)(M − 1) + j . Obviously, designs that achieve the lower bound are E(s 2 )-optimal.

15.4.3 BIBD-Based Supersaturated Designs Interesting relationships between certain BIB designs and supersaturated designs were established by Nguyen (1996), Cheng (1997), and Liu and Zhang (2000). We shall not go into all the details here, but just describe one method based on the cyclic development of initial blocks as described by Liu and Zhang (2000). They establish the equivalence of existence of a supersaturated design with N = 2v, n = c(2v − 1) (with v ≥ 3, c ≥ 2 and when v is odd, c is even) and that of a BIB design with t = 2v − 1, b = c(2v − 1), r = c(v − 1), k = v − 1, λ = c(v − 2)/2, which is cyclically developed from c initial blocks.

602


Given a BIB design with parameters specified above and the treatments labeled as 0, 1, 2, . . . , t − 1 = 2v − 2, the procedure to generate a supersaturated design can be described as follows: 1. Develop the first initial block cyclically, giving rise to blocks 1, 2, . . . , 2v − 1. 2. Consider an array of 2v − 1 columns, each column containing the elements 0, 1, 2, . . . , 2v − 1 (in that order), which correspond to the 2v − 1 blocks obtained in step 1. In column j replace the v − 1 elements corresponding to the treatments in block j by +1 and the remaining v elements by −1. 3. Add a +1 in the first position to each column, thus generating the first 2v − 1 columns of the array (15.1). 4. Repeat steps 1, 2, and 3 for each of the remaining c − 1 initial blocks, thus generating altogether the n = c(2v − 1) columns of (15.1). We shall illustrate this procedure in the next example. Example 15.2 For v = 4, c = 2 the two initial blocks are given by Liu and Zhang (2000) as (0 1 3) and (0 1 5). The individual steps then are: (1) Develop (0 1 3) and (0 1 5) mod 7: 0 1 2 3 4 5 6 1 2 3 4 5 6 0 3 4 5 6 0 1 2 0 1 2 3 4 5 6 1 2 3 4 5 6 0 5 6 0 1 2 3 4 (2) + (3) + (4) and writing + for +1 and − for −1: + + + − + − − −

+ − + + − + − −

+ − − + + − + −

+ − − − + + − +

+ + − − − + + −

+ − + − − − + +

+ + − + − − − +

+ + + − − − + −

+ − + + − − − +

+ + − + + − − −

+ − + − + + − −

+ − − + − + + −

+ − − − + − + +

+ + − − − + − +

The columns represent the supersaturated design (15.1) with N = 8, n = 14.

603

THREE-LEVEL SUPERSATURATED DESIGNS

Liu and Zhang (2000) provide a list of initial blocks for constructing a large number of supersaturated designs with N = 2v and n = c(2v − 1) for 3 ≤ v ≤ 12. All these designs are E(s 2 ) optimal. 15.5 THREE-LEVEL SUPERSATURATED DESIGNS Clearly, from a practical point of view two-level supersaturated designs are the most important. But in some situations three-level designs may be called for if the few active factors are expected to exhibit linear and/or quadratic effects. Whereas there exist several methods of constructing two-level supersaturated designs, there exist only very few for constructing three-level designs (see Yamada and Lin, 1999; Yamada et al., 1999; Fang, Lin, and Ma, 2000). We shall describe briefly one method due to Yamada and Lin (1999). Let C = [c1 , c2 , . . . , cn∗ ] be a two-level supersaturated design or saturated main effect plan in N ∗ runs. Let φ ab (C) be an operator that changes the elements in C from −1 (or −) to a and +1 (or +) to b. Then consider the matrix   12 φ (C) φ 12 (C) φ 13 (C) φ 23 (C)   D = φ 23 (C) φ 13 (C) φ 23 (C) φ 12 (C)

(15.10)

φ 31 (C) φ 23 (C) φ 12 (C) φ 13 (C) = D1

D2

D3

D4

(15.11)

Array (15.10) represents a three-level supersaturated design with N = 3N ∗ runs and n = 4n∗ factors. Each level of each factor occurs N ∗ times, and 2n > N . For n = 2n∗ or n = 3n∗ a combination of subdesigns D i (i = 1, 2, 3, 4) from (15.11) can be used to form a suitable supersaturated design. We shall illustrate this with the following example. Example 15.3 Suppose we want to examine 14 three-level factors accounting for 28 effects, and suppose further that we are allowed 24 runs. A suitable design can be constructed by choosing for C the saturated main effect plan with N ∗ = 8 and n∗ = 7, given by the following Hadamard matrix  + −  −  + C= − +  + −

+ + − − + − + −

+ + + − − + − −

− + + + − − + −

+ − + + + − − −

− + − + + + − −

 − −  +  −  + +  + −

(15.12)

604


Using D 2 and D 3 from (15.11) and applying the transformations to (15.12) yields the supersaturated design D: 2 1 1 2 1 2 2 1 3 1 1 3 1 3 3 1 3 2 2 3 2 3 3 2

2 2 1 1 2 1 2 1 3 3 1 1 3 1 3 1 3 3 2 2 3 2 3 2

2 2 2 1 1 2 1 1 3 3 3 1 1 3 1 1 3 3 3 2 2 3 2 2

1 2 2 2 1 1 2 1 1 3 3 3 1 1 3 1 2 3 3 3 2 2 3 2

2 1 2 2 2 1 1 1 3 1 3 3 3 1 1 1 3 2 3 3 3 2 2 2

1 2 1 2 2 2 1 1 1 3 1 3 3 3 1 1 2 3 2 3 3 3 2 2

1 1 2 1 2 2 2 1 1 1 3 1 3 3 3 1 2 2 3 2 3 3 3 2

3 1 1 3 1 3 3 1 3 2 2 3 2 3 3 2 2 1 1 2 1 2 2 1

3 3 1 1 3 1 3 1 3 3 2 2 3 2 3 2 2 2 1 1 2 1 2 1

3 3 3 1 1 3 1 1 3 3 3 2 2 3 2 2 2 2 2 1 1 2 1 1

1 3 3 3 1 1 3 1 2 3 3 3 2 2 3 2 1 2 2 2 1 1 2 1

3 1 3 3 3 1 1 1 3 2 3 3 3 2 2 2 2 1 2 2 2 1 1 1

1 3 1 3 3 3 1 1 2 3 2 3 3 3 2 2 1 2 1 2 2 2 1 1

1 1 3 1 3 3 3 1 2 2 3 2 3 3 3 2 1 1 2 1 2 2 2 1

Using the max χ 2 criterion (15.4), Yamada and Lin (1999) show that if max | sij | = d, say, over all i, j (1 ≤ i < j ≤ n∗ ), then for the design D in (15.10) or a combination of subdesigns from (15.11) we have

(N + 9d)2 N max χ = max , 8N 2

2

This result is useful because it allows us to compare competing designs with regard to the maximum dependency between two factors, the smaller this value, the better the design.

15.6 ANALYSIS OF SUPERSATURATED EXPERIMENTS Because the number of runs, N , is smaller than the number of parameters to be estimated, (p − 1)n, we cannot apply the usual analysis to the data from a supersaturated experiment, that is, an experiment using a supersaturated design.

605

ANALYSIS OF SUPERSATURATED EXPERIMENTS

Therefore, the analysis of such data must be predicated on the principle (or assumption) of active factor sparsity (Box and Meyer, 1986a, 1986b) or, in the case of p-level designs, effect sparsity as each factor is associated with p − 1 effects. Furthermore, we assume that all interactions are negligible. To put this in the context of a linear regression model, we express the observation vector y = (y1 , y2 , . . . , yN ) as y = Xβ + e

(15.13)

where X is an N × [(p − 1)n + 1] known matrix, which is determined by the design matrix, β is the [(p − 1)n + 1] vector of unknown parameters, and e is the error vector with the variance–covariance structure for a completely randomized design (see I.6.3.4). Based on the assumption of effect sparsity we may rewrite (15.13) as β (15.14) y = (X 1 , X2 ) 1 + e β2 where β 1 represents the vector of nonzero or large effects (including the intercept) and β 2 represents the vector of zero or negligible effects. Our aim, obviously, is to identify β 1 , that is, the “true” model. This necessitates applying variable selection procedures. An obvious choice is stepwise variable selection. This method may, however, not always be the best choice. If the number of factors is very large, the probability is high that some of the inert factors may be selected (type I error). Also, mainly because of the nonorthogonality of the X matrix, it may not always select active factors (type II error). A thorough discussion of these difficulties has been provided by Westfall, Young, and Lin (1998). To overcome in particular type I error problems, they recommend using adjusted P values at each step of the stepwise procedure, using familywise error rates. The familywise error rate should not be higher than .50 for inclusion of a variable in the model. It is for these reasons that Abraham, Chipman, and Vijayan (1999) recommend the use of all-subsets regression. For obvious reasons, this method has to be implemented in such a way that only subsets up to a certain prespecified size (< N ) will be considered. An advantage of this method is that it may produce different competing models, which allow the experimenter to make a choice based perhaps on subject matter considerations. Using similar arguments against the use of stepwise variable selection, Li and Lin (2002) propose a penalized least-squares approach to the variable selection problem. Let us rewrite (15.13) as y i = x i β + ei

(i = 1, 2, . . . , N )

(15.15)

where x i is the vector of input variables. For example, for the two-level design (15.15) is yi = β0 + x1i β1 + x2i β2 + · · · + xni βn + ei

606


or for the three-level design we have 2 2 2 β11 + x2i β2 + x2i β22 + · · · + xni βn + xni βnn + ei yi = β0 + x1i β1 + x1i

A form of penalized least squares is defined as Q(β) =

(p−1)n N 1 (yi − x i β)2 + p λ N ( | βj | ) 2N

(15.16)

j =0

i=1

(Fan and Li, 2001; Li and Lin, 2002), where pλN (·) is a penalty function and λN is a tuning parameter, which is usually chosen by some data-driven approach. Since this method cannot be applied to the full model (15.15), Li and Lin (2002) recommend to proceed in two steps: Apply stepwise variable selection (or subset selection, we might add) in step 1, using a low threshold, that is, a high α value, and then, in step 2, apply (15.16) to the model obtained in step 1. More formally, in (15.16) β = β 1 and x i = x 1i , say, where x 1i conforms to β 1 . An appropriate penalty function for supersaturated experiments is the smoothly clipped absolute deviation (SCAD) penalty, whose first-order derivative is

(aλ − |β|)+ pλ (|β|) = λ I (|β| ≤ λ) + I (|β| > λ) (15.17) (a − 1)λ for |β| > 0 with a = 3.7, and pλ (0) = 0 (Fan and Li, 2001; Li and Lin, 2002). Suppose β 1 in (15.14) has been identified in step 1 above to consist of d terms, say β(1) , β(2) , . . . , β(d) . Let β (0) 1 be an initial value close to the true value β 1 , when β(j ) is not very close to zero. The penalty function (15.17) can be locally approximated by a quadratic function as  (0)   pλ |β(j ) |  pλ (|β(j ) |) ≈ (15.18) β  | β (0) |  (j ) (j )

(j ) = 0 if β (0) is close to zero. With the local approximation (15.18), the and β (j ) solution to the penalized least square can be obtained iteratively as (0) −1 (1) β 1 = X1 X1 + N β1 X 1 y λ

with

λ

β (0) = diag 1

(15.19)

 (0)   pλ |β(j ) |  

(0) | β(j )|



j =1,2,..., d

For more details about the ridge regression (15.19) and description of the properties of the penalized least-squares estimators we refer the reader to Li and

ANALYSIS OF SUPERSATURATED EXPERIMENTS

607

Lin (2002). See also Li and Lin (2003) and Holcomb, Montgomery, and Carlyle (2003) for a comparison of analysis procedures. Regardless of which type of analysis is used, special care has to be exercised to interpret the results. We always have to remember that supersaturated designs are used for screening purposes, and the primary objective is to eliminate a large portion of inactive factors from further consideration. The nature of the designs do not guarantee that we always find the right answer. In any case, smaller follow-up experiments will be needed.

C H A P T E R 16

Search Designs

16.1 INTRODUCTION AND RATIONALE The usefulness of factorial experiments is based on the premise that high-order interactions are negligible. In previous chapters we have shown how this premise can be used to construct useful and efficient factorial designs through confounding and fractionation. Extreme examples of the latter category are, of course, main effect plans and supersaturated designs. They are based on the severe and often nonsustainable assumption that all interactions are negligible. It is this point that leads to the question whether one can construct designs that (1) have a relatively small number of runs and (2) allow the exploration of a limited number of low-level interactions. The emphasis here is on “limited number,” because even for a 2n factorial there are n(n − 1)/2 two-factor and n(n − 1)(n − 2)/6 three-factor interactions and to investigate them all would necessitate a resolution VII design. In most cases this may lead to a prohibitively large number of runs. But how can we limit the number of interactions to be investigated since we do not know which of the interactions are possibly nonnegligible? Clearly, we need to make some assumptions to overcome this difficulty. Rather than specifying the particular interactions we would like to investigate, we specify the number of interactions to be investigated. In other words, we assume that a specified number, say k, 2- or 3-factor interactions are possibly nonnegligible. We would then like to construct a design that allows us to (1) identify these nonnegligible interactions and (2) estimate these effects in addition to the main effects. This idea was first proposed and investigated by Srivastava (1975) when he introduced the concepts of search design and search linear model. We shall describe in this chapter some of his results and subsequent developments on this topic. 16.2 DEFINITION OF SEARCH DESIGN Let us consider here the case of n factors, each at two levels. We know from Section 7.4.2 that the true response for each treatment combination can be Design and Analysis of Experiments. Volume 2: Advanced Experimental Design By Klaus Hinkelmann and Oscar Kempthorne ISBN 0-471-55177-5 Copyright  2005 John Wiley & Sons, Inc.

608

609

PROPERTIES OF SEARCH DESIGNS

expressed as a linear function of the mean, all main effects and all interactions. We can divide the 2n − 1 factorial effects into three categories: C1 : contains the effects we want to estimate in any case, such as the main effects; C2 : contains effects we are interested in, but we know that only a few of them are nonnegligible, such as 2-factor interactions, C2 (2), or 2- and 3-factor interactions, C2 (2, 3); C3 : contains the remaining effects (such as higher-order interactions), which we consider to be negligible. Based on the above scenario we can then express a set of N observations, given by the N × 1 vector y, in terms of the following linear model: y = A 1 ξ 1 + A2 ξ 2 + e

(16.1)

where A1 , A2 are known matrices of order N × ν1 and N × ν2 , respectively, ξ 1 is a ν1 × 1 vector of unknown parameters consisting of the mean µ and the effects in C1 , ξ 2 is a ν2 × 1 vector of unknown parameters pertaining to the effects in C2 , and e is the N × 1 vector of errors with E(e) = 0

var(e) = V

(16.2)

where V represents the variance–covariance matrix for e under the error structure of the completely randomized design (see Section I.6.3.5). We assume now that at most k of the effects in C2 are nonnegligible, where k is quite small compared to ν2 . Let D be the design that gave rise to the observations y in (16.1). If y and hence A1 and A2 are such that we can estimate ξ 1 and the k nonnegligible effects of ξ 2 , then D is said to be a search design of resolving (or revealing) power (ξ 1 ; ξ 2 , k) and the model (16.1) is said to be a search linear model (Srivastava, 1975). 16.3 PROPERTIES OF SEARCH DESIGNS 16.3.1 General Case In the words of Srivastava (1984), model (16.1) represents a supermodel from which we want to extract an individual model based on the above criteria. Obviously, a search design that accomplishes this task must have certain properties. More specifically, the search design must have a sufficiently large number of runs and it must include a particular subset of all possible runs. Consider a design D with N runs and suppose that D is composed of two parts or subdesigns, D1 with N1 runs, and D2 with N2 runs. Let the observation vector y in (16.1) be partitioned conformably as y = (y 1 , y 2 ), where y 1 is N1 × 1 and is associated with D1 , while y 2 is N2 × 1 and is associated with D2 . Very often D1 represents a main effect plan to which treatment combinations in the form of D2 are added to generate a search design. In essence, this is a process of dealiasing main effects from certain interactions.

610

SEARCH DESIGNS

A fundamental result that provides a characterization of the properties that such a design must possess is due to Srivastava (1975) and can be stated as follows: Theorem 16.1 (i) Given model (16.1), a necessary condition for a design D to be a search design of revealing power (ξ 1 ; ξ 2 , k) is that rank(A1 , A20 ) = ν1 + 2k

(16.3)

for every N × 2k submatrix A20 of A2 . (ii) In the noiseless case, that is, for e ≡ 0 in (16.1) and (16.2), the condition (16.3) is also sufficient. We shall comment briefly on the noiseless case. It implies, of course, that there is no experimental error and no observational error (see Section I.6.3.4). Such a situation does not occur in practice. Even so, the noiseless case is important with regard to the construction of search designs, for if the structure of designs D1 and D2 and hence the structure of A1 and A2 in (16.1) is such that the search and estimation problem cannot be solved in this case, then it certainly cannot be solved when noise is present. Moreover, in the noiseless case the observations y in (16.1) are nonstochastic and hence the search problem is deterministic. This implies that the nonnegligible effects in ξ 2 can be identified with probability one if D satisfies the condition (16.3) of Theorem 16.1 In the stochastic case such identification can be achieved only with a certain probability, usually less than one. The following theorems provide not only a tool to check whether for a given design D condition (16.3) is satisfied but can also be used to obtain more explicit conditions that a search design has to satisfy. Theorem 16.2 (Srivastava and Ghosh, 1977) Consider model (16.1) and suppose that rank(A1 ) = ν1 . Then condition (16.3) holds if and only if for every N × 2k submatrix A20 of A2 we have rank [A20 A20 − A20 A1 (A1 A1 )−1 A1 A20 ] = 2k

(16.4)

A similar result is given in the following theorem. Theorem 16.3 (Srivastava and Gupta, 1979) Consider model (16.1) and suppose that rank(A1 ) = ν1 . Let A11 be a ν1 × ν1 submatrix of A1 of rank ν1 . Rearrange the observation vector y and write (16.1) as ∗ y1 A11 A12 = ξ1 + ξ2 + e (16.5) y ∗2 A21 A22 where y ∗1 , A11 , and A12 have ν1 rows and y2∗ , A21 , and A22 have N − ν1 rows. A necessary and sufficient condition for (16.3) to be satisfied is that every set of

611


2k columns of Q = A22 − A21 A−1 11 A12

(16.6)

are linearly independent. Note that in Theorem 16.3 the vector y ∗1 is not necessarily equal to y 1 obtained from D1 and, hence, y ∗2 is not necessarily equal to y 2 obtained from D2 as described earlier. However, if D1 is a saturated main effect plan with N1 = ν1 = n + 1, then y ∗1 = y 1 and, hence, y ∗2 = y 2 . 16.3.2 Main Effect Plus One Plans As stated earlier, in the context of search designs we assume a priory that at most k effects in ξ 2 of (16.1) are nonnegligible, where k is generally small. The most important and tractable situation from both the theoretical and practical point of view is the case k = 1. In addition, the most obvious starting point is to consider the case where ξ 1 is the vector of the mean, µ, and the n main effects, say F1 , F2 , . . . , Fn , of the n factors F1 , F2 , . . . , Fn , each at two levels. Thus, D1 might be a main effects plan. The resulting search design D is then referred to as a main effect plus one plan, or MEP.1 for short, with resolution III.1. Let us denote a treatment combination by x = (x1 , x2 , . . . , xn ) with x = 0 or 1( = 1, 2, . . . , n). Then we denote the Ni treatment combinations of Di (i = 1, 2) by x ij = (xij 1 , xij 2 , . . . , xij n ) for j = 1, 2, . . . , Ni . Now consider the following D1 with N1 = n + 2 treatment combinations and given by its design matrix   In    (16.7) D1 =  I n  0n Recall from Section 7.4.2 that the true response of treatment combination x can be expressed as

1 a(x) = µ + (−1) i αi (1−xi ) E α 2 α or, if we define E ∗α = 12 E α , as a(x) = µ +

(−1) i αi (1−xi ) E ∗α

(16.8)

α

Taking y ∗1 to be the first n + 1 observations in D1 it follows from (16.7) and (16.8) that A11 in Theorem 16.3 can be written as the patterned matrix 1 In A11 = In 2I n − In In

612

SEARCH DESIGNS

with A−1 11 having the same pattern. Following lengthy but elementary arguments we can then derive the general form of Q in (16.6) such that the conditions of Theorem 16.3 are satisfied (see Srivastava and Gupta, 1979), leading to the following theorem. 1 Theorem 16.4 (Srivastava and Gupta, 1979) Let D = D D 2 be the design matrix for the design D, where D 1 is given by (16.7). Let D 21 and D 22 be any two submatrices of D 2 of order (N − ν1 ) × n1 and (N − ν1 ) × n2 , respectively, with n1 , n2 ≥ 2. Let nij be the number of ones in the ith row of D 2j (j = 1, 2) and let for an integer z φ(z) =

 z

if z is even

z − 1

if z is odd

Then a necessary and sufficient condition for D to be a search design, where at most one element of ξ 2 is nonnegligible, is that for every pair of distinct submatrices D 21 and D 22 there exists a row in D 2 (say, the ith row), such that φ(ni1 ) φ(ni2 ) = φ(n1 ) φ(n2 ) [where i may be different for different pairs (D 21 , D 22 )]. This provides a convenient and simple check on D 2 for D to be an MEP.1. Whereas for the MEP.1 discussed above the vector ξ 2 in (16.1) consists of all possible interactions in 2n factorial, it seems more reasonable to restrict ξ 2 to include only 2- and 3-factor interactions, such that ν2 = n2 + n3 . A search design for identifying at most one nonnegligible effect among those interactions has been referred to by Gupta (1990) as a weak MEP.1. The structure of such an MEP.1 has been considered by Ghosh (1981) and Ohnishi and Shirakura (1985), and we shall give a brief description of their results without going through all the derivations. Let D1 be given by its design matrix   D1 =  

In 0n In In − I n

   

(16.9)

where D1 consists of N1 = n + 2 treatment combinations and represents a main effects plan. Let y 1 denote the observation vector obtained from D1 and let y 1 = A∗1 ξ 1 + A∗2 ξ 2 + e1

(16.10)

613


with ξ 1 and ξ 2 as defined above. Using the parameterization (16.8) it follows that 

 A∗1 =  1

In



In

1

−In In In

 

(16.11)

− 2I n

Now the columns of A∗2 in (16.9) and of A2 in (16.1) are obtained by appropriate multiplication of the elements in the columns of A∗1 (and A1 ). For example, to obtain the column corresponding to Fi1 Fi2 in A∗2 (and in A2 ) we simply multiply the corresponding elements of the columns belonging to Fi1 and Fi2 . Since D1 is a main effects plan, the columns of A∗2 are linear combinations of the columns of A∗1 . This fact can be exploited to obtain conditions for D2 such that condition (16.3) of Theorem 16.1 is satisfied for k = 1. This is done by considering all possible types of pairs of 2- and/or 3-factor interactions such as (Fi1 i2 , Fi1 i3 ), (Fi1 i2 , Fi3 i4 ), (Fi1 i2 , Fi1 i3 i4 ), and so forth with distinct indices i1 , i2 , i3 , i4 . This leads to the following theorem. Theorem 16.5 (Ohnishi and Shirakura, 1985) Let ω(z ) denote the weight of the vector z, which is defined as the number of nonzero elements in z. Further, let D 1 be given by (16.9) with row vectors of weights 0, n − 1, and n. Then a necessary and sufficient condition on D2 with N2 treatment combinations, such that D = D1 + D2 is a search design, is that there exist treatment combinations (x1 , x2 , . . . , xn ) in D2 satisfying the following conditions for any distinct integers ik (k = 1, 2, 3, 4, 5, 6) belonging to the set {1, 2, . . . , n}: (1)

ω(xi1 , xi2 ) = 1

xi3 = 0

(2)

ω(xi1 , xi2 ) = 1

ω(xi3 , xi4 ) = 1

(3a)

ω(xi1 , xi2 ) ≥ 1

ω(xi3 , xi4 ) = 0

(3b)

ω(xi1 , xi2 ) = 0

ω(xi3 , xi4 ) ≥ 1

(4a)

xi1 = 0, xi2 = 1

ω(xi3 , xi4 ) = 1

or

(4b)

ω(xi1 , xi2 ) ≥ 1

ω(xi3 , xi4 ) = 0

or

(4c)

ω(xi1 , xi2 ) = 0

ω(xi3 , xi4 ) = 2

(5a)

ω(xi1 , xi2 ) = 2

ω(xi3 , xi4 ) = 0

or

(5b)

ω(xi1 , xi2 ) = 0

ω(xi3 , xi4 ) = 2

or

(5c)

ω(xi1 , xi2 , xi3 , xi4 ) = 1

xi5 = 1

or

(5d)

ω(xi1 , xi2 , xi3 , xi4 ) = 3

xi5 = 0

or

614

SEARCH DESIGNS

(6a)

ω(xi1 , xi2 ) = 0

ω(xi3 , xi4 , xi5 ) ≥ 2 or

(6b)

ω(xi1 , xi2 ) ≥ 1

ω(xi3 , xi4 , xi5 ) ≤ 1

(7a)

ω(xi1 , xi2 , xi3 ) ≤ 1

ω(xi4 , xi5 , xi6 ) ≥ 2 or

(7b)

ω(xi1 , xi2 , xi3 ) ≥ 2

ω(xi4 , xi5 , xi6 ) ≤ 1

Even though the conditions in Theorem 16.5 are intended to provide a mechanism for checking condition (16.3) for all possible submatrices A20 , they themselves are not very easy to check since the number of cases to be considered is quite large. The conditions prove, however, to be useful in a sequential construction of search designs (see Section 16.4). 16.3.3 Resolution V Plus One Plans In the context of fractional factorial designs resolution V designs are of particular importance as long as the assumption of negligible interactions involving three or more factors is reasonable. If we suspect, however, that at most k of those interactions are nonnegligible, then a search design may be appropriate. Again, the case k = 1 is of interest. Such a search design is referred to as a resolution V.1 design. For resolution V designs the parameter vector ξ 1 in (16.1) contains the mean, the main effects, and the 2-factor interactions, with n ν1 = 1 + n + 2 components, and the vector of negligible effects, ξ 2 , consists of the remaining interactions with ν2 = 2n − ν1 terms. For purposes of a search design ξ 2 may be (i) as described above or (ii) restricted to involve only the 3-factor interactions, in which case ν2 = n3 . The case (i) above was considered by Srivastava and Ghosh (1976, 1977), whereas case (ii) was considered by Shirakura and Ohnishi (1985). In both cases the starting points are balanced resolution V plans (see Section 13.6.5). The reason for the restriction to these balanced arrays is the fact that it reduces greatly the number of cases one has to consider in order to verify whether condition (16.4) is indeed met. Exploiting the form of balanced arrays in light of the rank condition (16.4) leads then to the identification of those resolution V designs that are also resolution V.1 designs. We shall list a few examples in Section 16.4.2. 16.3.4 Other Search Designs In the previous section we have considered search designs for 2n factorials that allow us to search for one (k = 1) additional interaction among those interactions that are ordinarily assumed to be negligible. Extensions to k = 2 have been considered using Theorem 16.1 [for resolution III.2 plans see, e.g., Shirakura (1991)

LISTING OF SEARCH DESIGNS

615

and Chatterjee, Deng, and Lin (2001)]. Another extension is that to factors with more than two levels. Ghosh and Burns (2001), for example, consider search designs for the 3n factorial, whereas Chatterjee and Mukerjee (1986, 1993) consider the most general case for symmetrical and asymmetrical factorials for k = 1 and k ≥ 3, respectively. Another application of the principles of a search design is given by Ghosh and Burns (2002). They assume that ξ 1 in (16.1) contains only the mean and ξ 2 contains the main effects of the n factors. By assuming that at most k (< n) factors have nonnegligible effects, they are able to reduce the number of treatment combinations necessary to identify and estimate those effects compared to the number of treatment combinations in an orthogonal main effects plan. Note that this idea is similar to that used in supersaturated designs (see Section 15.1) except that the number of factors here is “small.”

16.4 LISTING OF SEARCH DESIGNS There are not very many methods of constructing search designs. The existing methods fall essentially into three categories: 1. Start with a “minimal” design D1 that allows estimating ξ 1 in (16.1) assuming that ξ 2 is negligible; then add treatment combinations in the form of D2 such that the final design D = D1 + D2 satisfies the conditions of a search design. 2. Start with a “larger” design D1 that allows estimating ξ 1 and show that D1 itself satisfies the conditions of a search design. 3. Start with an “intuitive” design D1 and show that it satisfies the conditions of a search design, or modify it accordingly. Methods 1 and 2 have been used for resolution III.1 and V.1, respectively, and we shall give some examples below. Method 3 is essentially a trial-and-error method and requires a great amount of intuition and insight. 16.4.1 Resolution III.1 Designs Starting with the design D1 given in (16.9), Ghosh (1981) suggested to construct the design D2 iteratively. More precisely, let D2,n be the design D2 for n factors with N2,n treatment combinations, then D2,n+1 can be represented by the design matrix D 2,n z1,n+1 (16.12) D 2,n+1 = D ∗2,n z2,n+1 where D 2,n+1 is a N2,n+1 × (n + 1) matrix, D ∗2,n is a (N2,n+1 − N2,n ) × n matrix, and zn+1 = (z1,n+1, , z2,n+1 ) is an N2,n+1 × 1 vector. All the elements

616

SEARCH DESIGNS

in D 2,n+1 are, of course, either 0 or 1. We note here that it is possible that N2,n+1 = N2,n . To proceed iteratively has the advantage that it greatly reduces the number of cases that have to be considered to check the conditions of Theorem 16.5. Using this method, Ohnishi and Shirakura (1985) have provided designs D2,n for 3 ≤ n ≤ 8, which are given in Table 16.1 with a summary of the various model and design parameters as given in Table 16.2. It illustrates the point that the total number N of treatment combinations required is relatively small but increases proportionately, of course, as n, the number of factors, increases. 16.4.2 Resolution V.1 Designs In order to reduce the computational load of checking the rank condition (16.4), construction of resolution V.1 designs has been limited to balanced arrays (see Section 13.6.5). Srivastava and Ghosh (1977) and Shirakura and Ohnishi (1985) provide tables of such search designs for n = 4, 5, 6, 7, 8, and we present some examples of their results in Table 16.3. Table 16.1 Design Matrices D 2, n for 3 ≤ n ≤ 8 Factors n

N2, n

1

2

3

4

5

6

7

8

1 2

1 0

0 0

0 1

1 1

0 0

0 0

1 1

0 1

3 4 5

1 0 0

0 1 0

0 0 1

0 0 0

1 1 1

0 0 1

1 1 0

1 1 0

6 7 8

1 1 1

0 0 0

0 0 0

1 0 1

0 1 1

1 1 0

0 0 0

1 1 0

9

1

0

0

0

1

1

0

0

10

0

1

0

0

1

0

1

0

Table 16.2 Number of Parameters and Runs for Search Design of Resolution III.1 n

ν1

ν2

N1

N2

N

3 4 5 6 7 8

4 5 6 7 8 9

4 10 20 35 56 84

5 6 7 8 9 10

2 5 5 8 8 10

7 11 12 16 17 20

617

ANALYSIS OF SEARCH EXPERIMENTS

Table 16.3 Resolution V.1 Designs for 4 ≤ n ≤ 8 Factors Srivastava and Ghosh (1977) n

ν1

ν2

N

4 5 6 7 8

11 16 22 29 37

5 16 42 99 219

15 21 28 36 45

λ 1 1 1 1 1

1 1 1 1 1

1 1 1 1 1

1 0 0 0 0

0 1 0 0 0

0 10 010 0010

Ohnishi and Shirakura (1985) ν2

N

20 35 56

28 35 44

λ

1001011 01000110 010000110

Let Sn,j denote the set of all treatment combinations with exactly j factors at level 1 and n − j factors at level 0, and let λj denote the number of times that each treatment combination in Sn,j appears in the design (array) (j = 0, 1, . . ., n). Of course, each vector (treatment combination) in Sn,j has weight j and λ = (λ0 , λ1 , . . ., λn ) is referred to as the index set of the array. The resolution V.1 designs, which are also referred to as BARE 5.1 plans by Srivastava and Ghosh (1976), are then completely described by their index sets, as given in Table 16.3. The Srivastava–Ghosh designs contain n Nn = 1 + 2n + 2 runs and can be represented easily as D = Sn,0 ∪ Sn, 1 ∪ Sn,2 ∪ . . . ∪ Sn,n−1 . Both types of search designs given in Table 16.3 are minimal designs, but designs with a larger number of runs are given in the two references. We also note that restricting ξ 2 in (16.1) does not reduce the size of the design appreciably. 16.5 ANALYSIS OF SEARCH EXPERIMENTS 16.5.1 General Setup We shall call an experiment that employs a search design a search experiment. The analysis of data from a search experiment are based, of course, on submodels of (16.1), which we write as y = A1 ξ 1 + A21 (ζ )ζ + e

(16.13)

where ζ refers to k possible components of ξ 2 in (16.1) and A21 (ζ ) refers to the column vectors in A2 corresponding to ζ . The analysis will then be performed for each possible ζ from ξ 2 . For purposes of characterizing and comparing the outcomes from the various analyses we shall denote by ζ 0 the true (unknown) nonnegligible effects in ξ 2 . We shall then refer to model (16.13) with ζ = ζ 0 as M0, and with ζ = ζ 0 as M1, where M1 represents a series of competing models to M0.

618

SEARCH DESIGNS

16.5.2 Noiseless Case Even though in practice there does not exist a situation where e ≡ 0 in (16.3), we have pointed out earlier that this noiseless case is important for purposes of constructing search designs. The noiseless case is equally important and instructive for describing and comparing the various analyses based on M0 and M1. We shall illustrate this for the case k = 1 by first constructing a data set with known parameters ξ 1 and ξ 2 and then using (16.13) with e ≡ 0 to analyze those data. Example 16.1 Let us consider the case of n = 3 factors and the search design with D2 given in Table 16.1. Thus, in model (16.1),   µ A    ξ1 =   B  C

 AB  AC    ξ2 =    BC  ABC 

Let µ = 10, A = 4, B = −2, C = 3, AB = 2, AC = 0, BC = 0, ABC = 0. Then using (16.8) [see also (7.30)] with A∗ = 2, B ∗ = −1, C ∗ = 1.5, (AB)∗ = 1 we obtain the true yield y = a(x) for the seven treatment combinations of the design D as follows: Treatment Combination 1 0 0 1 1 1 0

1 0 1 0 1 0 0

1 0 1 1 0 0 1

µ 10

A∗ 2

B∗ −1

C∗ 1.5

(AB)∗ 1

y

1 1 1 1 1 1 1

−1 −1 −1 1 1 1 −1

−1 −1 1 −1 1 −1 −1

−1 −1 1 1 −1 −1 1

1 1 −1 −1 1 −1 −1

13.5 8.5 7.5 13.5 10.5 10.5 11.5

The analysis of these data, using SAS PROC GLM, is presented in Tables 16.4a–16.4c for the models with ζ = AB, AC, and BC, respectively. We shall comment briefly on some aspects of the output: 1. Since AB = ζ0 , the model used in Table 16.4 a provides a perfect fit to the data as indicated by the fact that SS(Error) = SSE(M0) = 0, which is equivalent to the fact that the estimated effects are equal to the true = A, and that the predicted values effects, for example, A y are equal to the observed values y. In other words, the true model is identified correctly.


Table 16.4 Data and Analysis for Search Experiment (Noiseless Case) options nodate pageno=1; data search; input A B C y; datalines; 1 1 1 13.5 0 0 0 8.5 0 1 1 7.5 1 0 1 13.5 1 1 0 10.5 1 0 0 10.5 0 0 1 11.5 ; run; proc print data=search; title1 'TABLE 16.4'; title2 'DATA FOR SEARCH EXPERIMENT'; title3 '(NOISELESS CASE)'; run; proc glm data=search; class A B C; model y=A B C A*B/e; estimate 'A' A -1 1; estimate 'B' B -1 1; estimate 'C' C -1 1 ; estimate 'A*B' A*B 1 -1 -1 1/divisor=2; output out=predicta p=yhat r=resid stdr=eresid; title1 'TABLE 16.4a'; title2 'ANALYSIS WITH MODEL Y=A B C A*B'; title3 '(NOISELESS CASE)'; run; proc print data=predicta; run; proc glm data=search; class A B C; model y=A B C A*C/e; estimate 'A' A -1 1; estimate 'B' B -1 1; estimate 'C' C -1 1 ; estimate 'A*C' A*C 1 -1 -1 1/divisor=2; output out=predictb p=yhat r=resid stdr=eresid; title1 'TABLE 16.4b';

619

620

SEARCH DESIGNS

Table 16.4 (Continued ) title2 'ANALYSIS WITH MODEL A B C A*C'; title3 '(NOISELESS CASE)'; run; proc print data=predictb; run; proc glm data=search; class A B C; model y=A B C B*C/e; estimate 'A' A -1 1; estimate 'B' B -1 1; estimate 'C' C -1 1; estimate 'B*C' B*C -1 1 1 -1/divisor=2; output out=predictc p=yhat r=resid stdr=eresid; title1 'TABLE 16.4c'; title2 'ANALYSIS WITH MODEL Y=A B C B*C'; title3 '(NOISELESS CASE)'; run; proc print data=predictc; run; Obs

A

B

C

1 2 3 4 5 6 7

1 0 0 1 1 1 0

1 0 1 0 1 0 0

1 0 1 1 0 0 1

y 13.5 8.5 7.5 13.5 10.5 10.5 11.5

------------------------------------------------A. ANALYSIS WITH MODEL Y=A B C A*B (NOISELESS CASE) The GLM Procedure Class Level Information Class

Levels

Values

A

2

0 1

B

2

0 1

C

2

0 1

621


Table 16.4 (Continued ) Number of observations Read Number of observations Used

7 7

General Form of Estimable Functions Effect

Coefficients

Intercept

L1

A A

0 1

L2 L1-L2

B B

0 1

L4 L1-L4

C C

0 1

L6 L1-L6

A*B A*B A*B A*B

0 0 1 1

0 1 0 1

L8 L2-L8 L4-L8 L1-L2-L4+L8

Dependent Variable: y

DF

Sum of Squares

Mean Square

F Value

Pr > F

Model

4

31.42857143

7.85714286

Infty

F

1 1 1 1

13.76190476 1.66666667 10.00000000 6.00000000

13.76190476 1.66666667 10.00000000 6.00000000

Infty Infty Infty Infty

F

Model

4

42.09222138

10.52305535

22.00

0.0439

Error

2

0.95654033

0.47827017

Corrected Total

6

43.04876171

Source

Source A B C A*B

R-Square

Coeff Var

Root MSE

y Mean

0.977780

6.217804

0.691571

11.12243

DF

Type I SS

Mean Square

F Value

Pr > F

1 1 1 1

16.20876430 4.81496682 15.73393923 5.33455104

16.20876430 4.81496682 15.73393923 5.33455104

33.89 10.07 32.90 11.15

0.0283 0.0866 0.0291 0.0792

627


Table 16.5 (Continued ) Source A B C A*B

DF

Type III SS

Mean Square

F Value

Pr > F

1 1 1 1

29.79504504 11.74600417 19.51928067 5.33455104

29.79504504 11.74600417 19.51928067 5.33455104

62.30 24.56 40.81 11.15

0.0157 0.0384 0.0236 0.0792

Estimate

Standard Error

t Value

Pr > |t|

4.45683333 -2.79833333 3.60733333 1.88583333

0.56466519 0.56466519 0.56466519 0.56466519

7.89 -4.96 6.39 3.34

0.0157 0.0384 0.0236 0.0792

Parameter A B C A*B Obs

A

B

C

y

yhat

resid

1 2 3 4 5 6 7

1 0 0 1 1 1 0

1 0 1 0 1 0 0

1 0 1 1 0 0 1

13.792 9.013 7.445 15.187 10.176 10.606 11.638

13.7877 8.5218 7.4450 14.7002 10.1803 11.0928 12.1292

0.00433 0.49117 -0.00000 0.48683 -0.00433 -0.48683 -0.49117

eresid 0.39928 0.39928 . 0.39928 0.39928 0.39928 0.39928

------------------------------------------------B. ANALYSIS WITH MODEL Y=A B C A*C (WITH N(0, 1) NOISE) The GLM Procedure Class Level Information Class

Levels

Values

A

2

01

B

2

01

C

2

01

Number of observations 7 Dependent Variable: y

DF

Sum of Squares

Model

4

Error

2

Corrected Total

6

43.04876171

Source

Mean Square

F Value

Pr > F

39.22872871

9.80718218

5.13

0.1696

3.82003300

1.91001650

628

SEARCH DESIGNS

Table 16.5

(Continued ) R-Square

Coeff Var

Root MSE

y Mean

0.911263

12.42564

1.382033

11.12243

Source A B C A*C Source A B C A*C

DF

Type I SS

Mean Square

F Value

Pr > F

1 1 1 1

16.20876430 4.81496682 15.73393923 2.47105837

16.20876430 4.81496682 15.73393923 2.47105837

8.49 2.52 8.24 1.29

0.1004 0.2533 0.1030 0.3733

DF

Type III SS

Mean Square

F Value

Pr > F

1 1 1 1

20.14284038 6.03605400 11.88633750 2.47105837

20.14284038 6.03605400 11.88633750 2.47105837

10.55 3.16 6.22 1.29

0.0832 0.2174 0.1301 0.3733

Estimate

Standard Error

t Value

Pr > |t|

3.66450000 -2.00600000 2.81500000 1.28350000

1.12842560 1.12842560 1.12842560 1.12842560

3.25 -1.78 2.49 1.14

0.0832 0.2174 0.1301 0.3733

Parameter A B C A*C Obs

A

B

C

y

yhat

resid

eresid

1 2 3 4 5 6 7

1 0 0 1 1 1 0

1 0 1 0 1 0 0

1 0 1 1 0 0 1

13.792 9.013 7.445 15.187 10.176 10.606 11.638

13.4865 9.0130 8.5385 15.4925 9.3880 11.3940 10.5445

0.3055 0.0000 -1.0935 -0.3055 0.7880 -0.7880 1.0935

0.79792 . 0.79792 0.79792 0.79792 0.79792 0.79792

------------------------------------------------C. ANALYSIS WITH MODEL Y=A B C B*C (WITH N(0, 1) NOISE) The GLM Procedure Class Level Information Class

Levels

Values

A

2

01

B

2

01

C

2

01

629


Table 16.5 (Continued ) Number of observations 7 Dependent Variable: y

DF

Sum of Squares

Mean Square

F Value

Pr > F

Model

4

37.33955238

9.33488810

3.27

0.2477

Error

2

5.70920933

2.85460467

Corrected Total

6

43.04876171

Source

R-Square

Coeff Var

Root MSE

y Mean

0.867378

15.19055

1.689558

11.12243

Source A B C B*C Source A B C B*C

DF

Type I SS

Mean Square

F Value

Pr > F

1 1 1 1

16.20876430 4.81496682 15.73393923 0.58188204

16.20876430 4.81496682 15.73393923 0.58188204

5.68 1.69 5.51 0.20

0.1400 0.3236 0.1434 0.6959

DF

Type III SS

Mean Square

F Value

Pr > F

1 1 1 1

21.99952017 7.07094704 13.32209004 0.58188204

21.99952017 7.07094704 13.32209004 0.58188204

7.71 2.48 4.67 0.20

0.1090 0.2562 0.1633 0.6959

Estimate

Standard Error

t Value

Pr > |t|

3.82966667 -2.17116667 2.98016667 0.62283333

1.37951795 1.37951795 1.37951795 1.37951795

2.78 -1.57 2.16 0.45

0.1090 0.2562 0.1633 0.6959

Parameter A B C B*C Obs

A

B

C

y

yhat

resid

1 2 3 4 5 6 7

1 0 0 1 1 1 0

1 0 1 0 1 0 0

1 0 1 1 0 0 1

13.792 9.013 7.445 15.187 10.176 10.606 11.638

12.5333 7.8947 8.7037 15.3273 10.1760 11.7243 11.4977

1.25867 1.11833 -1.25867 -0.14033 0.00000 -1.11833 0.14033

eresid 0.97547 0.97547 0.97547 0.97547 . 0.97547 0.97547

630

SEARCH DESIGNS

The individual analyses for different models are given in Tables 16.5a–16.5c. The summary below shows that in this case we would have selected the true model over the others: Model

SS(Error)

M0:

A, B, C, AB

0.478

M1:

A, B, C, AC

1.910

M1:

A, B, C, BC

2.855

16.6 SEARCH PROBABILITIES The discussion of the preceding section can be expressed in terms of probabilities as follows: For the noiseless case we have P [SSE(M0) < SSE(M1) | M0, M1, V = 0] = 1 and for the noisy case we have P [SSE(M0) < SSE(M1) | M0, M1, V = 0] < 1

(16.14)

Let us assume, for operational purposes, that var(e) = V = σe2 I . If σe2 = ∞, then M0 and M1 are indistinguishable, and hence we can rewrite (16.14) more precisely as P [SSE(M0) < SSE(M1) | M0, M1, σe2 = ∞] =

1 2

The probability P [SSE(M0) < SSEM1) | M0, M1, σe2 ] is called a search probability, which assumes values between 12 and 1 (Ghosh and Teschmacher, 2002). The expression in (16.14) can be written out more explicitly as follows. To simplify the notation, let us write ξ1 θ= A(ζ ) = [A1 , A21 (ζ )] M(ζ ) = A(ζ ) A(ζ ) ζ Then the error sum of squares, SSE, for model (16.13), which we shall denote more precisely by s(ζ )2 , can be written as s(ζ )2 = y [I − Q(ζ )]y where

(16.15)

Q(ζ ) = A(ζ ) [M(ζ )]− A(ζ )

(see Section I.4.4.3). Rather than using (16.15) in (16.14) and minimizing s(ζ )2 , Shirakura, Takahashi, and Srivastava (1996) propose maximizing h(ζ , y) = y [Q(ζ ) − Q]y

631

SEARCH PROBABILITIES

where

Q = A1 (A1 A1 )−1 A1

and

s(ζ )2 = y (I − Q)y − h(ζ , y)

They then define P =

min min P [h(ζ 0 , y) > h(ζ , y)] ζ 0 ⊂ ξ 2 ζ ∈ A(ξ 2 ; ζ )

(16.16)

as the search probability for a given search design, where A(ξ z ; ζ ) denotes the set of all possible ζ of ξ 2 , of which at least one effect is not of ζ 0 , and ζ 0 represents the vector of the nonnegligible effects in ξ 2 . For the case k = 1 and assuming that e ∼ N (0, I σe2 ), Shirakura, Takahashi, and Srivastava (1996) derive an explicit expression for P (ζ0 , ζ, σe2 ) = P [h(ζ0 , y) − h(ζ, y)]

(16.17)

after rewriting (16.13) as y = A1 ξ 1 + a(ζ )ζ + e

(16.18)

where ζ is one effect of ξ 2 and a(ζ ) is the column vector in A2 corresponding to ζ , and ζ0 represents the nonnegligible effect in ξ 2 . The expression for (16.17) is then given as follows (see Ghosh and Teschmacher, 2002): P (ζ0 , ζ, σe2 ) = .5 + 2[φ(ρc1 (ζ0 , ζ )) − .5]

× [φ(ρ r(ζ0 , ζ0 ) − c12 (ζ0 , ζ )) − .5]

(16.19)

where φ(·) is the standard normal cdf, r(γ , η) = a (γ )[I − A1 (A1 A1 )−1 A1 ]a(η) √ for (γ , η) = (ζ0 , ζ0 ), (ζ0 , ζ ), (ζ, ζ ), x(ζ0 , ζ ) = r(ζ0 , ζ )/ r(ζ0 , ζ0 ) r(ζ, ζ ) , c12 (ζ0 , ζ ) = [r(ζ0 , ζ0 )/2][1 − | x(ζ0 , ζ ) |], and ρ = | ζ0 |/σe . It follows immediately from (16.19) that .5 ≤ P (ζ0 , ζ, σe2 ) ≤ 1, as stated before, and that P (ζ0 , ζ, σe2 ) increases as ρ increases. Using the search design and parameters in Example 16.2 we illustrate the use of (16.19) in the next example. Example 16.3 From Example 16.1 we obtain for use in (16.18) 1 −1 1 a(ζ0 ) = a (AB) = 1 1 −1 −1 1 −1 −1 −1 a(ζ1 ) = a (AC) = 1 1 −1 1 −1 −1 1 −1 a(ζ2 ) = a (BC) = 1 1

632

and

SEARCH DESIGNS

       A1 =       

1



1

1

1

 −1 −1 −1   −1 1 1    1 −1 1    1 1 −1    1 −1 −1 

1

−1 −1

1

1 1 1 1

1

with r(ζ0 , ζ0 ) = 6, r(ζ0 , ζ1 ) = 3, r(ζ0 , ζ2 ) = −1, r(ζ1 , ζ1 ) = 1.5, rζ2 , ζ2 ) = 5.5, x(ζ0 , ζ1 ) = 1, x(ζ0 , ζ2 ) = −0.1741, c12 (ζ0 , ζ1 ) = 0, c12 (ζ0 , ζ2 ) = 2.4777, ρ = 1. We then obtain P (ζ0 , ζ1 , 1) = .5, P (ζ0 , ζ2 , 1) = .9144. Individually, the results show that there is a much higher probability of distinguishing between the true model (including AB) and the model including BC. If we adopt (16.16) as the overall measure for the given design and parameters, then the search probability P is only .5. Ghosh and Teschmacher (2002) show how search probabilities can be used to compare different search designs. They develop criteria so that such comparisons can be made without requiring a specific value of ρ = | ζ0 |/σe .

C H A P T E R 17

Robust-Design Experiments

As we have mentioned before, factorial designs in their various forms were originally developed and used in the context of agricultural experimentation. They have since found wide application in almost all fields of empirical research. Important contributions to factorial experimentation has come from applications in engineering and manufacturing, a development usually associated with the name Taguchi. The so-called Taguchi method and variations of this method are concerned with quality engineering and product improvement. According to Taguchi (1993, p. 4), “the objective of quality engineering is to discover the countermeasure which results in an overall reduction in the total loss to both the company and society. It is therefore important to evaluate the two kinds of losses—the loss due to function and one due to harmful effects.” Thus, the basic philosophy here rests on the concept of societal loss due to imperfect production and “environmental” factors and the role that experimental design can play to reduce that loss. We shall discuss some of the underlying ideas in the following sections.

17.1 OFF-LINE QUALITY CONTROL An important aspect of any production process is quality control. Typically this is performed during the manufacturing phase of the process. Various forms of inspection sampling can be applied to ensure that the product meets certain specification and that the process is in control (see, e.g., Montgomery, 1991). There exists, however, another form of quality control. It is conducted at the product or process design stage, thus preceding the manufacturing stage. This type of quality control is therefore referred to as off-line quality control. Its aim is to improve product manufacturability and reliability, and to reduce product development and lifetime costs (Kackar, 1985). Design and Analysis of Experiments. Volume 2: Advanced Experimental Design By Klaus Hinkelmann and Oscar Kempthorne ISBN 0-471-55177-5 Copyright  2005 John Wiley & Sons, Inc.

633

634

ROBUST-DESIGN EXPERIMENTS

It is in this context that Taguchi (1986, 1987) introduced the application of what he called robust parameter design (parameter here refers to the characteristic of a product). According to Taguchi (1986, p. 98) its purpose “is to adjust the parameter levels so that the objective characteristic will not vary much even if the system and environmental parameters change. It is a search for the parameter levels at which the characteristic is stable.” More explicitly, Montgomery (1999) reformulates this in terms of making (1) products and processes insensitive to environmental factors or other factors hard to control and to variation transmitted from components and (2) finding levels of the process variables that achieve a mean performance at a specified (target) value and at the same time reducing variability around that value.

17.2 DESIGN AND NOISE FACTORS With the previous statements we have implicitly acknowledged the existence of two categories of factors that affect the characteristic of the product: Design factors and noise factors. This categorization is a new feature that we have not encountered before, at least not in this explicit form. It is, however, an important feature in conducting and analyzing robust-design experiments. Design factors, sometimes also referred to as design parameters (Kackar, 1985) or control factors (Taguchi, 1993), are those factors that can be controlled easily during the manufacturing process, the values of which remain essentially fixed at “optimal” levels. Noise factors, on the other hand, are factors that are hard to control, such as environmental factors, and, as a consequence, their levels will vary within a certain range during the production process and/or during subsequent use of the product [for a detailed description of various types of noise factors see, e.g., Wu and Hamada (2000)]. As an example we mention an experiment performed in a semiconductor factory to determine the influence of five factors on the transistor gain for a particular device (Montgomery, 1999). The three factors implant dose, drivein time, and vacuum level are easy to control and hence represent the design factors. The factors oxide thickness and temperature are hard to control during the manufacturing process and hence represent the noise factors. The underlying idea of the Taguchi method is to incorporate the noise factors into the product design phase by varying their settings systematically over a range representative of the variation during typical manufacturing and/or use conditions. The aim then is to find settings for the design factors so that the influence of the noise factors on the product performance is minimized or, to express it differently, to make the product robust with respect to the noise factors. It is for this reason that this methodology is referred to as robust design or robust parameter design (or parameter design for short).

635

MEASURING LOSS

17.3 MEASURING LOSS As mentioned earlier, the notion of loss to society and to the company forms the underpinning of the Taguchi method. To assess the loss we need to specify a loss function as a function of a value of the performance characteristic of the product, say Y , and the target value, say τ , for that performance characteristic. We can think of many loss functions, but in practice it is usually difficult to come up with the actual loss function. However, in many situations an appropriate loss function is given by (Y ) = c(Y − τ )2

(17.1)

where c is some constant, which can be determined if (Y ) is known for some value of Y . One way to determine c in (17.1) was described by Kackar (1985). Let (τ − , τ + ) be the consumer’s tolerance interval in the sense that if Y falls outside this interval, then the product needs either to be repaired or discarded. Suppose the cost for that action is A dollars. It follows then from (17.1) that A = c2 and hence c = A/2 . The loss function (17.1) is symmetric around τ . This may not always be appropriate as the loss may depend on whether Y < τ or Y > τ . In this case a simple variation of (17.1) may lead to the form (Y ) =

c1 (Y − τ )2

if Y < τ

c2 (Y − τ )2

if Y > τ

(17.2)

where c1 and c2 are constants that can be determined using arguments similar to those given above. A special case of (17.2) occurs if τ = 0. Then (17.2) reduces to (Y ) = c2 Y 2

if Y > 0

(17.3)

In either of the cases (17.1), (17.2), (17.3) we consider the expected loss L = E[(Y )]

(17.4)

as the criterion to assess the quality of a product or process. The expectation in (17.4) is taken with respect of the distribution of Y during the product’s lifetime and across different users. The latter is represented to a great extent by some or all of the noise factors.

636


17.4 ROBUST-DESIGN EXPERIMENTS In order to minimize the average loss (17.4) we perform an experiment, the robust-design experiment. The purpose of this experiment is to determine the optimal levels for the control factors in conjunction with reducing the product’s performance around the target value. There are several ways in which such an experiment can be performed making use of some of the factorial designs we have discussed in previous chapters or certain types of response surface designs (see I.12.3 and I.12.4). 17.4.1 Kronecker Product Arrays Distinguishing between the two sets of factors, design and noise factors, suggests the use of the following experimental design: Choose an appropriate fractional factorial design for the design factors and another one for the noise factors and then combine these two designs in the form of a Kronecker product design; that is, each design factor level combination occurs with each noise factor level combination. Taguchi (1986) referred to these two component designs as inner array and outer array, respectively. Moreover, he suggested the use of main effect plans in the form of orthogonal arrays. As an illustration consider the following example. Example 17.1 Suppose we have four design factors A1 , A2 , A3 , A4 , each at three levels, and three noise factors, B1 , B2 , B3 , each at two levels. The inner array is then given by OA(9, 4, 3, 2; 1) and the outer array by OA(4, 3, 2, 2; 1) (see Example 13.9). Then the Kronecker product design or the so-called cross array consists of 9 × 4 = 36 runs. An important aspect of the cross array is that the interactions between individual design and noise factors can be estimated (see Example 13.9). If such interactions exist, they can be exploited to choose the design factor settings such that the product variation over the noise factor settings will be reduced (we shall return to this point in Section 17.5). On the other hand, the number of runs required for a cross array may be quite large, even when we use minimal designs for both the design and noise factors. This disadvantage of cross arrays will become even more pronounced if we allow at least some 2-factor interactions between design factors by using a resolution IV or resolution V design instead of a main effect plan. 17.4.2 Single Arrays The last two comments above provide a strong argument to look for alternative experimental designs in connection with robust-design experiments. This can be achieved by combining the design and noise factors into one set of factors and then obtain an appropriate fractional factorial design that achieves the objectives of the experiment. Such experimental designs are referred to as single array or


637

common array. We shall illustrate this approach with some examples. For a listing of useful single arrays see Wu and Hamada (2000). Example 17.2 Suppose we have three design factors A1 , A2 , A3 and two noise factors B1 , B2 , each at two levels. The usual Taguchi method would cross 2 the 23−1 III array with the 2 array with 4 × 4 = 16 runs. Suppose, however, that we are interested also in estimating 2-factor interactions between design factors. In that case the crossed array would be of the form 23 × 22 with 8 × 4 = 32 array based on the runs. A single array can be obtained by considering the 25−1 V identity relationship I = A1 A2 A3 B1 B2 with only 16 runs. Example 17.3 Suppose we have five design factors A1 , A2 , A3 , A4 , A5 and two noise factors B1 , B2 , each at two levels. The usual cross array would be of 7−2 2 the form 25−2 III × 2 with 8 × 4 = 32 runs. The single array 2IV , also with 32 runs, based on the identity relationship I = A1 A2 A3 A5 = A1 A2 A4 B1 B2 = A3 A4 A5 B1 B2 has the additional property that it allows us to estimate at least some of the 2-factor interactions among design factors, assuming that only interactions involving three or more factors are negligible. Example 17.4 Suppose we have three design factors A1 , A2 , A3 , each at three levels, and two noise factors, B1 , B2 , each at two levels. The usual cross 3−1 2 array would be of the form 33−1 III × 2 with 9 × 4 = 36 runs. Here the 3III , referred to L9 (33 ) by Taguchi, is based on I = A1 A2 A23 . An alternative single array in the form of a modified central composite design for five factors (see I.12.4.3) was proposed by Myers, Khuri, and Vining (1992). Their design consists of three parts: (1) The factorial part is a 25−1 based on I = A1 A2 A3 B1 B2 and V using only the high and low levels of each factor; (2) the axial part consists of the axial points for the control factors, keeping the noise factors at their intermediate level (assuming that the noise factors are quantitative factors), and (3) n0 center runs. Using the response surface methodology convention of denoting the low, intermediate and high levels by −1, 0, 1, respectively, the final design is given in Table 17.1. This design, consisting of 16 + 6 + n0 = 22 + n0 runs, allows us to estimate all main effects, all 2-factor interactions between design factors, and the 2-factor interactions between design and noise factors. The single arrays discussed above provide some insight into their general structure. It also shows that single arrays may be more economical than crossed arrays or that they provide more information. These are aspects that may determine which type of array we should choose for a given situation. However, to some extent this choice will also be influenced by the type of modeling we intend to use for data from such experiments, that is, how we intend to use the data in order to arrive at a robust product design. We shall discuss some of these issues in the following sections.

638


Table 17.1 Central Composite Design for Three Design Factors and Two Noise Factors A1

A2

A3

B1

B2

−1 −1 −1 −1 −1 −1 −1 1 1 1 −1 1 −1 1 1 −1 1 1 −1 1 −1 1 1 1 −1 1 −1 −1 1 1 1 −1 1 −1 1 1 −1 1 1 −1 1 1 −1 −1 1 1 1 −1 1 −1 1 1 1 −1 −1 −1 −1 −1 −1 1 −1 −1 −1 1 −1 −1 −1 1 −1 −1 −1 1 −1 −1 −1 1 −1 −1 −1 −1 -----------------------------------−1 0 0 0 0 1 0 0 0 0 0 −1 0 0 0 0 1 0 0 0 0 0 −1 0 0 0 0 1 0 0 -----------------------------------0

0

0

0

0

n0

17.5 MODELING OF DATA Just as there are different approaches to deciding on a statistical design for a robust-design experiment, there are different ways of analyzing the data from such an experiment. The proposal by Taguchi (1986) to use various forms of signal-to-noise ratios (S/N ratios) to analyze and interpret such data has come under a certain amount of criticism (e.g., Box, 1988) and as a consequence other methods of looking at the data from these experiments have been proposed. We shall indicate some of these ideas in the following sections. 17.5.1 Location and Dispersion Modeling As we have mentioned earlier, the novel idea associated with robust-design experiments is not only to identify the factors that affect the average response, but also to identify the factors that contribute to the variability of the response. In

639

MODELING OF DATA

the practical setting of product design this means more specifically that we need to model both the location and the dispersion of the response or, as it is often referred to, the performance characteristic in order to find an appropriate level combination of the control factors that will achieve the desired value of or target for the performance characteristic. Let us consider the cross array of Section 17.4.1. Suppose the inner or design array D1 consists of N1 treatment combinations and the outer or noise array D2 consists of N2 treatment combinations. The total number of runs in D = D1 × D2 then is N = N1 N2 . Let A denote the design factor settings, then, following Nair (1986), we can write a location-scale model for the (possibly transformed) observations as Y = µ(A) + σ (A)e

(17.5)

where the random component e does not depend on A and is assumed to have mean 0 and variance 1. The terms µ(A) and σ (A) represent the location and dispersion effects, respectively. We should emphasize that using model (17.5) implies that the experiment has been carried out according to a completely randomized design. In practice this may not always be possible and some form of blocking may have to be used as discussed in Chapters 11 and 12. Another possibility is the use of a split-plottype design (see I.13.4), where, for example, the treatment combinations of the design array represent the whole-plot treatments and the treatment combinations of the noise array represent the split-plot treatments. In either case it is easy to modify model (17.5) by either incorporating the block effects or the whole-plot error term, both of which are independent of the design factors. Using the loss function (17.1) the expected loss (17.4) for the response (17.5) can be written as L = E[(Y )] = c E(Y − τ )2 = c[µ(A) − τ ]2 + c σ 2 (A)

(17.6)

This form of L shows explicitly that in order to minimize the mean-squared error (17.6) we need to identify the setting of the design factors such that µ(A) achieves the target value τ and σ 2 (A) is minimized. Now let A = (A1 , A2 , . . . , An ) be the set of design factors employed in the experiment. Then the aim of the subsequent analysis is to divide A into three sets, say, A = (A1 , A2 , A3 ) (Phadke, 1982), where A1 represents the factors that affect dispersion (they are referred to as control factors); A2 represents the factors that affect location but not dispersion (they are referred to as adjustment factors); A3 represents the factors that neither affect location nor dispersion.

640


According to this division the factors included in A1 and A2 represent the actual design factors (Box, 1988). Such a division is possible only if we can get rid of dispersion effects that arise because of a dependence of σ on µ (Box and Meyer, 1986b; Box, 1988). This can often be achieved through an appropriate transformation of the original data or through the use of a performance measure independent of adjustment (PerMIA) as proposed by Leon, Shoemaker, and Kackar (1987). We assume that Y in (17.5) is of that form. To achieve the division of A let us now consider the observations Y from the experiment based on the cross array D. More specifically, let yij be the (possibly transformed) observation for the ith treatment combination of D1 and the j th treatment combination of D2 (i = 1, 2, . . . , N1 ; j = 1, 2, . . . , N2 ). In some sense the N2 observations, stemming from D2 , for each treatment combination of D1 can be considered as replications for that treatment combination. It is, therefore, plausible to analyze

y i. =

N2 1 yij N2

and si2 =

j =1

N2 1 (yij − y i. )2 N2 − 1 j =1

for i = 1, 2, . . . , N1 , separately as measures of location and variability, respectively. The analysis of the y i. conforms, of course, to the usual analysis of fractional factorial designs. The si2 are obtained over the same design points of D2 and, hence, any differences among them can be attributed to the design factors. This provides justification for analyzing the si2 , or more appropriately log si2 , in the same way as the y i. in order to identify the factors belonging to A1 and A2 . For both variables we can use the tools of the analysis of variance. We shall illustrate this approach with the data obtained from a cross-array experiment described by Engel (1992). Example 17.5 (Engel, 1992) Seven design factors and three noise factors, each at two levels, were included in the experiment in an effort to identify the design factors that had an effect on the mean of and variation in the percentage of shrinkage of products made by injection moulding. The factors are (rather than using the labels A1 , A2 , . . . , A7 and B1 , B2 , B3 we retain the labels used by Engel): Design Factors

Noise Factors

A: cycle time B: mould temperature C: cavity thickness D: holding pressure E: injection speed F: holding time G: gate size

M: percentage regrind N: moisture content O: ambient temperature

641

MODELING OF DATA

3−1 The cross-array D is given by crossing D1 = 27−4 III with D2 = 2III , were D1 has the identity relationship

I = ABC = DEFG = ABCDEFG = ADE = BCDE = AFG = BCFG = BEG = ACEG = BDF = ACDF and D2 has the identity relationship I = MNO The 32 design points with their observations are listed in Table 17.2 together with the values for si2 (note that as a matter of convenience each si2 value is listed four times for the four replications of each design point in D1 ). The results of the analysis of variance for the y i. and the log si2 are given in Table 17.2. Concerning the location modeling aspect, we may conclude by looking informally and somewhat subjectively at the P values (or half-normal plots) for the estimated main effects, that the factors A, D, and possibly G constitute A2 and hence represent the adjustment factors. A separate regression analysis using A, D, and G yields the prediction equation = 2.3375 + 0.85xA − 0.5625xD − 0.4625x6 Y where xA , xD , xG represent the coded levels of A, D, and G, respectively. Together with engineering and production considerations this equation can then be used to adjust the process so that a particular target, τ , can be achieved. With regard to the dispersion modeling the results point clearly to factor F as having a pronounced effect on the variability and hence constituting the only control factor. A look at the data shows, indeed, that the si2 for those treatment combinations in D1 with F = 0 are much smaller than for those with F = 1. For a different approach to dispersion modeling and analysis of the same data set we refer the reader to Engel (1992). 17.5.2 Dual-Response Modeling We have pointed out earlier that cross-array experiments allow for the estimation of 2-factor interactions between design and noise factors. Yet, in the preceding discussion we have assumed that such interactions either do not exist or have been eliminated through an appropriate transformation of the original observations. Also, the location and dispersion modeling approach as discussed in the previous section is obviously not appropriate for single-array experiments. In order to deal with these situations we shall describe now a two-step modeling procedure that Myers, Khuri, and Vining (1992) and Vining and Myers (1990) referred to as a dual-response approach. The basic idea is to express an

642


Table 17.2 Robust-Design Experiment Cross Array Data Points, Data, and Analysis options nodate pageno=1 data robust; input A B C D E F G M N logS=log(S); datalines; 0 0 0 0 0 0 0 0 0 0 2.2 0 0 0 0 0 0 0 1 0 1 2.3 0 0 0 1 1 1 1 0 0 0 0.3 0 0 0 1 1 1 1 1 0 1 2.7 0 1 1 0 0 1 1 0 0 0 0.5 0 1 1 0 0 1 1 1 0 1 0.4 0 1 1 1 1 0 0 0 0 0 2.0 0 1 1 1 1 0 0 1 0 1 1.8 1 0 1 0 1 0 1 0 0 0 3.0 1 0 1 0 1 0 1 1 0 1 3.0 1 0 1 1 0 1 0 0 0 0 2.1 1 0 1 1 0 1 0 1 0 1 1.0 1 1 0 0 1 1 0 0 0 0 4.0 1 1 0 0 1 1 0 1 0 1 4.6 1 1 0 1 0 0 1 0 0 0 2.0 1 1 0 1 0 0 1 1 0 1 1.9 ; run;

LS=75; O Y S @@;

0.0092 0.0092 1.7700 1.7700 2.1000 2.1000 0.0092 0.0092 0.0025 0.0025 1.8733 1.8733 1.7625 1.7625 0.0067 0.0067

0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1

0 0 0 0 1 1 1 1 0 0 0 0 1 1 1 1

0 0 0 0 1 1 1 1 1 1 1 1 0 0 0 0

0 0 1 1 0 0 1 1 0 0 1 1 0 0 1 1

0 0 1 1 0 0 1 1 1 1 0 0 1 1 0 0

0 0 1 1 1 1 0 0 0 0 1 1 1 1 0 0

0 0 1 1 1 1 0 0 1 1 0 0 0 0 1 1

0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0

2.1 2.3 2.5 0.3 3.1 2.8 1.9 2.0 3.1 3.0 4.2 3.1 1.9 2.2 1.9 1.8

0.0092 0.0092 1.7700 1.7700 2.1000 2.1000 0.0092 0.0092 0.0025 0.0025 1.8733 1.8733 1.7625 1.7625 0.0067 0.0067

proc print data=robust; title1 'TABLE 17.2'; title2 'ROBUST-DESIGN EXPERIMENT'; title3 'CROSS ARRAY DATA POINTS AND DATA'; run; proc glm data=robust; class A B C D E F G; model Y logS = A B C D E F G; estimate 'A' A -1 1/divisor=16; estimate 'B' B -1 1/divisor=16; estimate 'C' C -1 1/divisor=16; estimate 'D' D -1 1/divisor=16; estimate 'E' E -1 1/divisor=16; estimate 'F' F -1 1/divisor=16; estimate 'G' G -1 1/divisor=16; title2 'IDENTIFICATION OF ADJUSTMENT AND DISPERSION FACTORS'; title3 'THROUGH ANALYSIS OF VARIANCE'; run; proc reg data=robust; model Y = A D G; title2 'ESTIMATING PREDICTION EQUATION'; title3 'USING ADJUSTMENT FACTORS A, D, G'; run; quit;

643

MODELING OF DATA

Table 17.2 (Continued ) Obs

A

B

C

D

E

F

G

M

N

O

Y

S

LogS

1

0

0

0

0

0

0

0

0

0

0

2.2

0.0092

-4.68855

2

0

0

0

0

0

0

0

0

1

1

2.1

0.0092

-4.68855

3

0

0

0

0

0

0

0

1

0

1

2.3

0.0092

-4.68855

4

0

0

0

0

0

0

0

1

1

0

2.3

0.0092

-4.68855

5

0

0

0

1

1

1

1

0

0

0

0.3

1.7700

0.57098

6

0

0

0

1

1

1

1

0

1

1

2.5

1.7700

0.57098

7

0

0

0

1

1

1

1

1

0

1

2.7

1.7700

0.57098

8

0

0

0

1

1

1

1

1

1

0

0.3

1.7700

0.57098

9

0

1

1

0

0

1

1

0

0

0

0.5

2.1000

0.74194

10

0

1

1

0

0

1

1

0

1

1

3.1

2.1000

0.74194

11

0

1

1

0

0

1

1

1

0

1

0.4

2.1000

0.74194

12

0

1

1

0

0

1

1

1

1

0

2.8

2.1000

0.74194

13

0

1

1

1

1

0

0

0

0

0

2.0

0.0092

-4.68855

14

0

1

1

1

1

0

0

0

1

1

1.9

0.0092

-4.68855

15

0

1

1

1

1

0

0

1

0

1

1.8

0.0092

-4.68855

16

0

1

1

1

1

0

0

1

1

0

2.0

0.0092

-4.68855

17

1

0

1

0

1

0

1

0

0

0

3.0

0.0025

-5.99146

18

1

0

1

0

1

0

1

0

1

1

3.1

0.0025

-5.99146

19

1

0

1

0

1

0

1

1

0

1

3.0

0.0025

-5.99146

20

1

0

1

0

1

0

1

1

1

0

3.0

0.0025

-5.99146

21

1

0

1

1

0

1

0

0

0

0

2.1

1.8733

0.62770

22

1

0

1

1

0

1

0

0

1

1

4.2

1.8733

0.62770

23

1

0

1

1

0

1

0

1

0

1

1.0

1.8733

0.62770

24

1

0

1

1

0

1

0

1

1

0

3.1

1.8733

0.62770

25

1

1

0

0

1

1

0

0

0

0

4.0

1.7625

0.56673

26

1

1

0

0

1

1

0

0

1

1

1.9

1.7625

0.56673

27

1

1

0

0

1

1

0

1

0

1

4.6

1.7625

0.56673

28

1

1

0

0

1

1

0

1

1

0

2.2

1.7625

0.56673

29

1

1

0

1

0

0

1

0

0

0

2.0

0.0067

-5.00565

30

1

1

0

1

0

0

1

0

1

1

1.9

0.0067

-5.00565

31

1

1

0

1

0

0

1

1

0

1

1.9

0.0067

-5.00565

32

1

1

0

1

0

0

1

1

1

0

1.8

0.0067

-5.00565

644


Table 17.2 (Continued ) IDENTIFICATION OF ADJUSTMENT AND DISPERSION FACTORS THROUGH ANALYSIS OF VARIANCE The GLM Procedure Dependent Variable: Y DF

Sum of Squares

Mean Square

F Value

Pr > F

Model

7

11.00000000

1.57142857

1.67

0.1646

Error

24

22.60000000

0.94166667

Corrected Total

31

33.60000000

Source

Source A B C D E F G Source A B C D E F G

Parameter A B C D E F G

R-Square

Coeff Var

Root MSE

y Mean

0.327381

43.12867

0.970395

2.250000

DF

Type I SS

Mean Square

F Value

Pr > F

1 1 1 1 1 1 1

5.78000000 0.18000000 0.12500000 2.53125000 0.66125000 0.01125000 1.71125000

5.78000000 0.18000000 0.12500000 2.53125000 0.66125000 0.01125000 1.71125000

6.14 0.19 0.13 2.69 0.70 0.01 1.82

0.0207 0.6659 0.7188 0.1141 0.4103 0.9139 0.1902

DF

Type III SS

Mean Square

F Value

Pr > F

1 1 1 1 1 1 1

5.78000000 0.18000000 0.12500000 2.53125000 0.66125000 0.01125000 1.71125000

5.78000000 0.18000000 0.12500000 2.53125000 0.66125000 0.01125000 1.71125000

6.14 0.19 0.13 2.69 0.70 0.01 1.82

0.0207 0.6659 0.7188 0.1141 0.4103 0.9139 0.1902

Estimate

Standard Error

t Value

Pr > |t|

0.05312500 -0.00937500 0.00781250 -0.03515625 0.01796875 -0.00234375 -0.02890625

0.02144291 0.02144291 0.02144291 0.02144291 0.02144291 0.02144291 0.02144291

2.48 -0.44 0.36 -1.64 0.84 -0.11 -1.35

0.0207 0.6659 0.7188 0.1141 0.4103 0.9139 0.1902

645

MODELING OF DATA

Table 17.2 (Continued ) IDENTIFICATION OF ADJUSTMENT AND DISPERSION FACTORS THROUGH ANALYSIS OF VARIANCE The GLM Procedure Dependent Variable: LogS DF

Sum of Squares

Mean Square

F Value

Pr > F

Model

7

266.4310915

38.0615845

Infty

F

1 1 1 1 1 1 1

1.5111751 0.6003957 0.2841756 0.3835367 0.7414489 261.7830683 1.1272911

1.5111751 0.6003957 0.2841756 0.3835367 0.7414489 261.7830683 1.1272911

Infty Infty Infty Infty Infty Infty Infty

|t|

1 2 3 4

22.3375000 23.1958333 16.2458333 16.7708333

0.5394137 0.5394137 0.5394137 0.5394137

2 The results of Kunert and Stufken (2002) are closely tied to the notions of balanced block designs and generalized Youden designs (see Kiefer, 1958, 1975a, 1975b). Definition 19.9 A block design with t treatments and blocks of size k is called a balanced block design if (1) each treatment is replicated equally often, (2) each pair of treatments occurs together the same number of times, and (3) |nij − k/t| < 1 for all i, j (i = 1, 2, . . . , t; j = 1, 2, . . . , b), where nij is the number of times treatment i appears in block j (see Section 1.3.1). Definition 19.10 A row–column design is called a generalized Youden design if it is balanced block design when considering rows as blocks and columns as blocks. Investigating the properties of the information matrix C τ under model (19.15) with regard to uniform optimality Kunert and Stufken (2002) introduced the following definition. Definition 19.11 A design d in t,n,p is called totally balanced if the following conditions hold: 1. 2. 3. 4.

d is a generalized Youden design; d is a balanced block design in the carryover effects; d is balanced for carryover effects; The number of subjects where both treatment u and u appear [p/t] + 1 times and treatment u does not appear in the last period is the same for every pair u = u .

To establish universal optimality we then have the following theorem. Theorem 19.10 (Kunert and Stufken, 2002) For t ≥ 3 and 3 ≤ p ≤ 2t, if a totally balanced design d ∗ in t,n,p exists, then d ∗ is universally optimal for the estimation of direct effects over t,n,p . No general methods of constructing such designs exist, but they can sometimes be obtained by modifying existing designs considered in earlier sections. We illustrate this with the following example. Example 19.10 For t = 3, p = 4, and n = 6 the extra-period design  1 3 d= 2 2

2 1 3 3

3 2 1 1

1 2 3 3

2 3 1 1

 3 1  2 2

713

COMMENTS ON OTHER MODELS

is universally optimal under model (19.1). Now d is strongly balanced for residual effects, rather than balanced. Hence d is not totally balanced. Changing the arrangement in the last period leads to the following totally balanced design: 

1 3 ∗ d = 2 1

2 1 3 2

3 2 1 3

1 2 3 1

2 3 1 2

 3 1  2 3

Thus d ∗ is universally optimal under model (19.15).

Kunert and Stufken (2002) point out the interesting fact that in optimal designs under model (19.15) generally no self carryover effects occur. This shows that optimality may not always be the ultimate criterion because in some applications it may be useful to obtain estimates of self carryover effects, for example, in connection with a single drug therapy for a chronic disease (Afsarinejad and Hedayat, 2002). 19.8.7 Factorial Treatment Structure So far we have tacitly assumed that we are dealing with t qualitative treatments. It is, however, entirely possible that we could have quantitative treatments. For example, our treatments may be represented by one or more factors at several levels. This fact may be used to construct special designs that achieve certain statistically desirable properties, such as certain types of orthogonality. Such designs were discussed by for example, Berenblut (1967, 1968) and Mason and Hinkelmann (1971). Here we shall discuss briefly yet another type of treatment structure, namely a factorial treatment structure. We shall limit ourselves to the case of 2m factorials, although the approach can be extended easily to other factorials. An obvious way of proceeding is to simply equate the 2m treatment combinations to the t treatments and use the designs we have discussed earlier. This will lead, however, very quickly to a prohibitively large number of periods. To reduce the number of periods we shall make the assumption that some interactions are negligible and then use appropriate systems of confounding. In order to use this approach it is useful to use in model (19.1) the reparameterization (7.42) and rewrite the model as yij = µ + πi + βj + a x d(i,j ) + ar zd(i−1,j ) + eij where x d(i,j ) represents the treatment combination x assigned by d to subject j in period i, and zd(i−1,j ) the treatment combination z assigned to subject j in period i − 1. Further, a(x d(i,j ) ) and ar (zd(i−1,j ) ) represent the true direct and residual effects in terms of main effect and interaction effect components α1 α2 αm α Eαα x = (Aα1 1 Aα2 2 · · · Aαmm )α x and Er,α x = (A1 A2 · · · Am )r,α x , respectively.

714

CROSSOVER DESIGNS

The method to construct suitable designs is to first construct the block design confounding selected interactions and their generalized interactions with blocks. These blocks constitute the generating subjects. Each generating subject is then developed into a Williams square as described in Section 19.5.1. Formally we can describe the result as in the following theorem. Theorem 19.11 (Shing and Hinkelmann, 1992) For a 2m factorial with = effects E β 1 , E β 2 , . . . , E β negligible (this set is made up of q independent interactions and their generalized interactions), where m ≥ q + 1, there exists a CO(2m , 2m , 2m−q ) such that all nonnegligible direct and residual effects and interactions are estimable. To construct such a design we assign 2m−q treatment combinations to each of q 2 subjects in arbitrary order by confounding E β 1 , E β 2 , . . . , E β with subjects. Each of the 2q initial subjects is then developed into a Williams square. We adjoin the 2q Williams squares to obtain the final CO(2m , 2m , 2m−q ). 2q−1

We illustrate this procedure in the following example. Example 19.11 Suppose we have a 23 factorial with factors A, B, C. Suppose further that we want to use p = 4 = 22 periods, that is, we have m = 3, q = 1. Thus we need to confound one interaction with subjects, say ABC. The final design is then given in Table 19.4, with subjects 1 and 5 being the initial subjects, and using the fact that the Williams square, that is, CO(4, 4, 4) is of the form 1 4 2 3

2 1 3 4

3 2 4 1

4 3 1 2

implying that for subject 1 we have the association (0, 0, 0) ≡ 1, (1, 0, 1) ≡ 2, (0, 1, 1) ≡ 3, (1, 1, 0) ≡ 4, with a similar association for subject 5. From the design in Table 19.4 we can estimate A, B, AB, C, AC, BC and Ar , Br , (AB)r , Cr , (AC)r , (BC)r . We point out, however, that the effects like Table 19.4 CO(23 , 8, 4) with ABC Confounded Subjects Periods 1 2 3 4

1

2

3

4

5

6

7

8

(0,0,0) (1,1,0) (1,0,1) (0,1,1)

(1,0,1) (0,0,0) (0,1,1) (1,1,0)

(0,1,1) (1,0,1) (1,1,0) (0,0,0)

(1,1,0) (0,1,1) (0,0,0) (1,0,1)

(1,1,1) (0,0,1) (0,1,0) (1,0,0)

(0,1,0) (1,1,1) (1,0,0) (0,0,1)

(1,0,0) (0,1,0) (0,0,1) (1,1,1)

(0,0,1) (1,0,0) (1,1,1) (0,1,0)

COMMENTS ON OTHER MODELS

715

Ar have no particular meaning, except that they represent linear combinations of residual effects. Obviously, for large values of m this method may lead to rather large designs. In that case we may need to use fractional factorials together with a system of confounding. More specifically, if we use a 2m−s fraction and 2m−s−q periods, then we can construct a CO(2m−s , 2m−s , 2m−s−q ) by the method described above. However, this must be considered carefully because this may not lead to practically useful designs for all values of m, s, and q. For another approach to constructing crossover designs with a factorial treatment structure we refer to Fletcher, Lewis, and Matthews (1990). They consider in particular the case of two factors with up to four levels in three or four periods. Their designs are obtained by cyclic generation from one or more suitably chosen initial sequences. A table of such sequences is provided by Fletcher, Lewis, and Matthews (1990).

APPENDIX A

Fields and Galois Fields

The notion of a finite field, in particular a Galois field, plays an important role in the construction of experimental designs (see, e.g., Chapters 3 and 13). We shall give here the elementary notions of fields and Galois fields. For a more thorough discussion of this topic the reader is referred to, for example, Carmichael (1956). Consider a set of s distinct elements (marks) U = {u0 , u1 , . . . , us−1 }. Definition A.1 The set U is said to be a finite field of order s if its elements satisfy the following conditions: 1. The elements may be combined by addition with the laws ui + uj = uj + ui ui + (uj + uk ) = (ui + uj ) + uk 2. The sum of any two elements is an element in U. 3. Given two elements ui and uk , there is a unique element uj such that ui + uj = uk . Then uj is said to be determined by subtraction, and we write uj = uk − ui . 4. The element having the additive property of zero is u0 so that u0 + uj = uj

or

u0 = uj − uj

for all j . 5. The elements may be combined by multiplication with the laws ui uj = uj ui ui (uj uk ) = (ui uj )uk ui (uj + uk ) = ui uj + ui uk Design and Analysis of Experiments. Volume 2: Advanced Experimental Design By Klaus Hinkelmann and Oscar Kempthorne ISBN 0-471-55177-5 Copyright  2005 John Wiley & Sons, Inc.

716

717

APPENDIX A

Table A.1 Primitive Marks of Finite Fields of Order p Order of Field (p) 3 5 7 11 13 17 19 23

Primitive Mark (u) 2 2 3 2 2 3 2 5

6. The product of any two elements is an element in U. 7. From the relations u0 ui = ui u0 = ui (uj − uj ) = ui uj − ui uj = u0 it follows that u0 has the multiplicative properties of zero. 8. Given any ui (= u0 ) and any uk , there is a unique uj such that ui uj = uk Then uj is said to be determined by division, and uj is called the quotient of uk by ui . 9. The quotient ui /ui has the multiplicative properties of unity and is chosen to be ui . The finite field of s = p elements, where p is a prime number, may be represented by u0 = 0, u1 = 1, u2 = 2, . . . , up−1 = p − 1. Addition and multiplication are the ordinary arithmetic operations, except that the resulting number is reduced mod p, that is, the resulting number is replaced by the remainder after dividing by p. Every element u ∈ U(u = u0 ) satisfies the equation up−1 = 1. If for a given u, p − 1 is the smallest power satisfying this equation, then u is called a primitive mark of the field. Then the marks u, u2 , u3 , . . . , up−1 are all distinct and hence represent, in some order, the elements of U. Examples of primitive marks (roots) for different values of p are given in Table A.1. As an illustration consider the following example. Example A.1 Consider p = 7, u = 3, u1 = 3, u2 = 32 = 2, u3 = 2 · 3 = 6, u4 = 6 · 3 = 4, u5 = 4 · 3 = 5, u6 = 5 · 3 = 1. The field of order p described above is a special case of a finite field of order s = p n , where p is a prime number. Such a field is referred to as a Galois field

718

APPENDIX A

and is generally denoted by GF(p n ). For n = 1 the elements are the residues modulo p as described above. For n > 1 the elements are obtained as follows. Let P (x) be a given polynomial in x of degree n with integral coefficients and let F (x) be any polynomial in x with integral coefficients. Then F (x) may be expressed as F (x) = f (x) + p · q(x) + P (x) Q(x) where f (x) = b0 + b1 x + b2 x 2 + · · · + bn−1 x n−1 and coefficients bi (i = 0, 1, . . . , n − 1) belonging to GF(p), that is, the set U = (0, 1, 2, . . . , p − 1). This relationship may be written as F (x) = f (x)

mod p, P (x)

(A.1)

and we say that f (x) is the residue modulis p and P (x). The functions F (x), which satisfy (A.1) when f (x), p, and P (x) are kept fixed, form a class. If p and P (x) are kept fixed, but f (x) is varied, p n classes may be formed since each coefficient in f (x) may take the p values 0, 1, 2, . . . , p − 1. It may be readily verified that the classes defined by the f (x)’s make up a field. For example, if Fi (x) belongs to the class corresponding to fi (x), and Fj (x) corresponding to fj (x), then Fi (x) + Fj (x) belongs to the class corresponding to fi (x) + fj (x). The other operations defined for a field are, likewise, satisfied, any function obtained by ordinary algebraic operations being replaced by its residue modulis p and P (x). In order that division be unique, it is also necessary that p be a prime number and that P (x) cannot be expressed in the form P (x) = P1 (x) P2 (x) + p P3 (x)

(A.2)

where Pi (x) are polynomials in x with integral coefficients, and the degrees of P1 (x) and P2 (x) being positive and less than the degree of P (x). Any P (x) that does not admit the representation is thus irreducible mod p. Summarizing the preceding discussion, we have the following definition. Definition A.2 For p a prime number and P (x) an irreducible polynomial mod p, the residue classes given by (A.1) form a finite field of order p n . This field is called a Galois field, and is denoted by GF(p n ). As an illustration we consider the following example. Example A.2 Let p = 3 and consider GF(32 ). Then P (x) is of degree 2, that is, of the form a0 + a1 x + a2 x 2 with a0 , a1 , a2 ∈ {0, 1, 2}. For P (x) to be

719

APPENDIX A

irreducible mod 3 we can exclude a0 = 0. Suppose we take a0 = 1. Then P (x) is of the form 1 + a1 x + a2 x 2 . If a1 = 0, a2 = 2 we have P (x) = 1 + 2x 2 , which is 0 for x = 1 and can be written as (1 + x)(1 + 2x) − 3x, that is, it is of the form (A.2), and hence is not irreducible mod 3. We then consider a1 = 0, a2 = 1, that is, P (x) = 1 + x 2 . For x = 0, 1, 2, P (x) takes the values 1, 2, 2 mod 3, respectively, and hence is irreducible mod 3. For P (x) = 1 + x 2 we then consider the possible f (x)’s, which are of the form b0 + b1 x. Hence, the elements of the GF(32 ) are: u0 = 0, u1 = 1, u1 = 2, u3 = x, u4 = 2x, u5 = 1 + x, u6 = 2 + x, u7 = 1 + 2x, u8 = 2 + 2x. Just as for the GF(p), there exists as primitive mark u such that the powers uj (j = 1, 2, . . . , pn − 1) represent the nonzero elements of GF(p n ). In the above case, u = 1 + x is such a primitive mark, since u1 = 1 + x = u5 u2 = 1 + 2x + x 2 = 2x = u4 u3 = 2x(1 + x) = 2x + 2x 2 = 2x + 1 + 2(1 + x 2 ) = 1 + 2x = u7 u4 = (1 + 2x)(1 + x) = 1 + 3x + 2x 2 = 2 + 2(1 + x 2 ) = 2 = u2 u5 = 2(1 + x) = 2 + 2x = u8 u6 = (2 + 2x)(1 + x) = 2 + 4x + 2x 2 = 4x = x = u3 u7 = x(1 + x) = x + x 2 = 2 + x + (1 + x 2 ) = 2 + x = u6 u8 = (2 + x)(1 + x) = 2 + 3x + x 2 = 1 = u1

The irreducible polynomials P (x) are not unique. It can be shown, however, that there always exists a particular irreducible P (x) such that u = x is a primitive element. Such a P (x) is called a primitive polynomial. A partial list of primitive polynomials is given in Table A.2 [for a more extensive list see Hedayat, Sloane, and Stufken (1999)]. To perform arithmetic operations in a GF it is convenient to first obtain addition and multiplication tables in terms of the powers of the primitive element, x, using the primitive polynomial (see Table A.2). We illustrate this with the following example. Example A.3 Consider the GF(22 ). The elements are u0 = 0, u1 = 1, u2 = x, u3 = 1 + x = x 2 , with P (x) = 1 + x + x 2 . We then have the following addition table:

u0 u1 u2 u3

u0

u1

u2

u3

u0 u1 u2 u3

u1 u0 u3 u2

u2 u3 u0 u1

u3 u2 u1 u0

720

APPENDIX A

Table A.2 List of Primitive Polynomials for Selected Values of p and n GF

P (x) 1 + x + x2 1 + x + x3 1 + x + x4 1 + x2 + x5 2 + x + x2 1 + 2x + x 3 2 + x + x2

22 23 24 25 32 33 52

or

0 x0 x1 x2

0

x0

x1

x2

0 x0 x1 x2

x0 0 x2 x1

x1 x2 0 x0

x2 x1 x0 0

For example, x 1 + x 2 = x(1 + x) = x 3 = 1 = x 0 Similarly, we obtain the following multiplication table:

u0 u1 u2 u3

u0

u1

u2

u3

u0 u0 u0 u0

u0 u1 u2 u3

u0 u2 u3 u1

u0 u3 u1 u2

0

x0

x1

x2

0 0 0 0

0 x0 x1 x2

0 x1 x2 x0

0 x2 x0 x1

or

0 x0 x1 x2 For example,

x2 · x2 = x4 = x3 · x = x = x1 With these tables we can then carry out all arithmetic operations in GF(22 ).

APPENDIX B

Finite Geometries

A finite geometry consists of a finite set of elements, called points, that are subject to certain conditions about points and lines. Such a geometry can be represented in terms of the marks of a Galois field, GF(p n ), where p is a prime. Of primary importance in the context of constructing experimental designs are K-dimensional projective and Euclidean geometries, denoted by PG(K, p n ) and EG(K, p n ), respectively.

PROJECTIVE GEOMETRY PG(K, P N ) A point of a K-dimensional projective geometry is defined by the ordered set of homogeneous coordinates (µ0 , µ1 , . . . , µK ) where µi ∈ GF(p n ) and at least one µi = u0 = 0. The symbol (µµ0 , µµ1 , . . ., µµK ) denotes the same point, where µ ∈ GF(p n ) and µ = u0 . The total number of points in the PG(K, p n ) then is p n(K+1) − 1 = 1 + p n + p 2n + · · · + p Kn pn − 1

(B.1)

since there exist p n(K+1) − 1 ordered sets = (0, 0, . . ., 0), and these may be arranged in groups of p n − 1 sets all of whose elements represent the same point. Now consider two distinct points (µ0 , . . ., µK ) and (ν0 , . . ., νK ). Then the line between these points is defined by the points (µµ0 + νν0 , µµ1 + νν1 , . . . , µµK + ννK )

(B.2)


721

722

APPENDIX B

where µ, ν ∈ GF(p n ), not both 0. The number of possible combinations of µ and ν is p 2n − 1, so that because of proportionality the line (B.2) contains p 2n − 1 = pn + 1 p n−1

(B.3)

points. The system of points and lines given by (B.1) and (B.3) is called a Kdimensional projective geometry PG(K, p n ). In (B.2) we have defined what we might call a 1-space in the PG(K, p n ). More generally, we can define an M-space (M < K) in PG(K, p n ) as follows: If the points (µi0 , µi1 , . . ., µiK ) (i = 0, 1, . . ., M) are any M + 1 points not in the same (M − 1)-space, then an M-space is defined by the points

M

µi µi0 ,

µi µi1 , . . . ,

µi µiK

(B.4)

i=0

where µ0 , µ1 , . . ., µM ∈ GF(p n ) and not all equal to 0. Alternatively, (B.4) can be described in terms of the solutions to K − M independent, linear, homogeneous equations ai0 µ0 + ai1 µ1 + · · · + aiK µK = 0

(B.5)

(i = 1, 2, . . ., K − M) with aij fixed. In order to find the number of M-spaces in the K-space (M < K), we need to find first the number of ways in which the M + 1 noncollinear base points can be chosen, such that they do not all lie in any (M − 1)-space: 1st point:

1 + p n + · · · + p Kn ≡ NK1

2nd point:

p n + · · · + p Kn ≡ NK2

3rd point: .. .

p 2n + · · · + p Kn ≡ NK3

(M + 1)th point:

p Mn + · · · + p Kn ≡ NK,M+1

and then the ways in which M + 1 points of a given PG(M, p n ) may be selected in a given order so that they do not all lie in any (M − 1)-space: 1st point

1 + p n + · · · + p Mn ≡ NM1

2nd point .. .

p n + · · · + p Mn ≡ NM2

(M + 1)th point

p Mn ≡ NM,M+1

723

APPENDIX B

Hence every PG(K, p n ) contains M+1

ψ(M, K, p n ) =

NKi

i=1 M+1

(B.6)

NMi

i=1

M-spaces, with NKi and NMi as defined above. Similarly, we can show that every PG(S, p n ) in PG(K, p n ) is contained in (S+1)n + · · · + p Kn × · · · × p Mn + · · · + p Kn p n ϕ(S, M, K, p ) = (S+1)n + · · · + p Mn × · · · × p (M−1)n + p Mn p Mn p (B.7) PG(M, p n )’s for S < M. In particular, for S = 0, every point is contained in ϕ(0, M, K, p n ) PG(M, p n ) in PG(K, p n ), and, for S = 1, every line is contained in ϕ(1, M, K, p n ) PG(M, p n ) in PG(K, p n ), or, since with every pair of points also the line containing them is in the PG(M, p n ), it follows that every pair of points is contained in ϕ(1, M, K, p n ) different PG(M, p n ). EUCLIDEAN GEOMETRY EG(K, P N ) Consider a PG(K, p n ) and the subset of points (µ0 , µ1 , . . ., µK ) for which µ0 = 0. Take µ0 = 1, then the points (1, µ1 , . . ., µK ) constitute an EG(K, p n ) consisting of p Kn points. The excluded points (0, µ1 , . . ., µK ) form a PG(K − 1, p n ) with homogeneous coordinates (µ1 , µ2 , . . ., µK ). Hence, the EG(K, p n ) is obtained by deleting from the PG(K, p n ) a PG(K − 1, p n ). Every PG(M, p n ) contained in PG(K, p n ) but not in PG(K − 1, p n ) becomes an EG(M, p n ) since by deleting PG(K − 1, p n ) from PG(K, p n ) we also delete PG(M − 1, p n ) from PG(M, p n ). The number of EG(M, p n ) contained in EG(K, p n ) then is ψ(M, K, p n )−ψ(M, K − 1, p n ), with ψ defined in (B.6). The number of EG(M, p n ) containing a given EG(S, p n ) with S < M is ϕ(S, M, K, p n ) as defined in (B.7).

APPENDIX C

Orthogonal and Balanced Arrays

ORTHOGONAL ARRAYS Definition C.1 A K × N matrix A, with entries from a set of s ≥ 2 elements, is called an orthogonal array of strength t, size N, K constraints and s levels if each t × N submatrix of A contains all possible t × 1 column vectors with the same frequency λ, denoted by OA[N, K, s, t; λ], where λ is called the index of A and N = λs t . For the case

= {0, 1, . . . , s − 1} consider the following example.

Example C.1 OA[9, 4, 3, 2; 1]: 

0  0 A=  0 0

0 1 1 1

0 2 2 2

1 0 1 2

1 1 2 0

1 2 0 1

2 0 2 1

2 1 0 2

 2 2   1  0

There are different methods of constructing orthogonal arrays [for a thorough discussion see Hedayat, Sloane, and Stufken (1999)]. We shall present here a particular method based on projective geometries. Theorem C.1 If we can find a matrix C = (cij ) of K rows and τ columns, whose elements cij belong to a GF(p n ), and for which every submatrix obtained by taking t rows is of rank t, then we can construct an OA[p nτ , K, p n , t; λ]. Proof Consider τ × 1 column vectors ξ whose coordinates belong to GF(p n ). Then there are p τ n different ξ i . Consider then the matrix A whose p τ n columns are the vectors C ξ i (i = 1, 2, . . ., p nτ ). A is then an orthogonal array: Design and Analysis of Experiments. Volume 2: Advanced Experimental Design By Klaus Hinkelmann and Oscar Kempthorne ISBN 0-471-55177-5 Copyright  2005 John Wiley & Sons, Inc.

724

725

APPENDIX C

Suppose A∗ is a t × p τ n submatrix of A, and C ∗ is the corresponding t × τ submatrix of C. Consider A∗ = C ∗ (ξ 1 , ξ 2 , . . . , ξ s ) with s = p τ n , and denote the ith column of A∗ by α i (i = 1, 2, . . ., s), that is, αi = C ∗ ξ i or c11 ξi1 + c12 ξi2 + · · · + c1τ ξiτ = αi1 c21 ξi1 + c22 ξi2 + · · · + c2τ ξiτ = αi2 .. . ct1 ξi1 + ct2 ξi2 + · · · + ctτ ξiτ = αit Assume without loss of generality that the first t columns of C ∗ are linearly independent. Then consider all the ξ i that have the same last τ − t components. There are p tn different such vectors. Each will give rise to a different α i . But the total number of different α i is p tn , that is, each possible α i is obtained exactly once from such a set. Now there are p n(τ −t) possible sets of vectors for which this holds. Hence each different α i is obtained exactly p n(τ −t) times. This implies that in A∗ each possible t × 1 column vector occurs λ = p n(τ −t) times and shows that A is an orthogonal array of strength t and index λ. The rows of C in Theorem C.1 may be interpreted as the coordinates of a PG(τ − 1, p n ) such that no t points lie in a space of dimension ≤ t − 2. To find a matrix C then means to find K points of a PG(τ − 1, p n ) with the above restriction. We illustrate this with the following example. Example C.2 Consider the case p n = 2. In the PG(τ − 1, 2) consider all points that are not in the (τ − 2)-space: x0 + x1 + · · · + xτ −1 = 0 There are exactly 2τ −1 such points, namely those with an odd number of unity components. No three of these points are on the same line, since in the PG(τ − 1, 2) each line has exactly three points, one of which is excluded. Hence we can construct an OA[2τ , 2τ −1 , 2, 3; λ]. Consider, for example, the case τ = 3: 

1  0 C=  0 1

0 1 0 1

 0 0   1  1

726

APPENDIX C



0 1 0 1  0 0 1 1 (ξ 1 , ξ 2 , . . . , ξ 8 ) = 0 0 0 0 and hence



0  0 A=  0 0

1 0 0 1

0 1 0 1

1 1 0 0

0 1 0 0 0 1 1 1 1

 1 1  1

0 0 1 1

 1 1   1  1

1 0 1 0

0 1 1 0

with λ = 1. It is easy to see that A is also an OA[2τ , 2τ −1 , 2, 2; 2].

BALANCED ARRAYS Balanced arrays, or B-arrays, were introduced by Chakravarti (1956) under the name partially balanced arrays. The following definition is adapted from Srivastava and Chopra (1973). Definition C.2 A K × N matrix B, with entries from a set of s ≥ 2 elements, is said to be a balanced array of strength t, size N , K constraints, and s levels if for each t × N submatrix B 0 of B and for each t × 1 vector x = (x1 , x2 , . . ., xt ) with elements from , we have λ(x, B 0 ) = λ(P (x), B 0 )

(C.1)

where λ(x, B 0 ) is the number of columns of B 0 that are identical with x, and P (x) is any vector obtained by permuting the elements of x. We denote such an array by BA[N, K, s, t; λ ], where λ is the vector of the totality of λ(x, B 0 ) as defined by Definition C.1. The term λ is referred to as the index set of the B-array. Note that if all elements of λ are identical, say equal to λ, then the B-array becomes an orthogonal array with index λ. For the case = {0, 1, 2, . . ., s − 1} we illustrate the notion of a B-array with the following example. Example C.3 Consider BA[15, 5, 3, 2; 1, 2]    B= 

0 0 1 2 1

0 1 0 2 2

0 2 2 0 1

0 1 2 1 0

0 2 1 1 2

1 1 2 0 2

1 2 1 0 0

1 0 0 1 2

1 2 0 2 1

1 0 2 2 0

2 2 0 1 0

2 0 2 1 1

2 1 1 2 0

2 0 1 0 2

2 1 0 0 1

    

727

APPENDIX C

where λ(x1 , x2 ) =

λ1 = 1

for x1 = x2

λ2 = 2

for x1 = x2

Condition (C.1) can be expressed operationally as follows. Let in x the symbol i occur ri times (0 ≤ i ≤ s − 1). Then s−1

ri = t

i=0

and each of the t!/r0 !r1 ! . . . rs−1 ! vectors obtained by permuting x has the same λ value. For the special case s = 2, the index set λ consists of t + 1 components, λ0 , λ1 , . . ., λt , corresponding to the number of nonzero components in x, which is also referred to as the weight of x. There exist several methods of constructing B-arrays [see, e.g., Chakravarti (1956, 1961), Srivastava and Chopra (1973), Aggarwal and Singh (1981)]. We shall mention here one method that is based on the existence of orthogonal arrays (Chakravarti, 1961). Suppose, an orthogonal array OA[N, K, s, t; λ] can be divided into two disjoint arrays such that one array is a B-array or a degenerate B-array [a degenerate B-array being one that has some but not all λ(x1 , x2 , . . . , xt ) equal to zero], with λ(x1 , x2 , . . . , xt ) < λ for all t × 1 vectors x. Then the remaining array is a B-array with λ parameters given by λ∗ (x1 , x2 , . . . , xt ) = λ − λ(x1 , x2 , . . . , xt ). As an illustration consider the following example. Example C.4 Consider OA[9, 3,  0 1 2  A= 0 1 2 0 1 2

3, 2; 1] 0 2 1

1 2 0 1 0 1 1 2 2 0 2 0

 2  0  1

Divide A into 

0  A1 =  0 0

 1 2  1 2  1 2



and

0  A2 =  2 1

1 2 0 0 1 1 2 0 2

 1 2  2 0  0 1

where A1 is a degenerate B-array and A2 then represents a BA[6, 3, 3, 2] with λ1 = 0, λ2 = 1, where λ1 , λ2 are as defined in Example C.3. In other situations we may be able to obtain a B-array by deleting not only columns but also rows from an orthogonal array (see Chakravarti, 1961).

APPENDIX D

Selected Asymmetrical Balanced Factorial Designs

1. 4 × 2 × 2, k = 4 (Li, 1944) Replication 1 Block

Replication 2 Block

Replication 3 Block

1

2

3

4

1

2

3

4

1

2

3

4

000 101 210 311

001 100 211 310

010 111 200 301

011 110 201 300

000 111 201 310

001 110 200 311

010 101 211 300

011 100 210 301

000 110 211 301

001 111 210 300

010 100 201 311

011 101 200 310

λ(000) = 3

h(000) =

2

λ(100) = 0

h(100) =

1 4

λ(010) = 0

h(010) =

0

λ(110) = 1

h(110) =

0

λ(001) = 0

h(001) =

0

λ(101) = 1

h(101) =

0

λ(011) = 0

h(011) =

λ(111) = 1

h(111) =

1 4 − 14

E(110) = E(101) = E(111) = 23 , all other E(z) = 1, E = 45 . Design and Analysis of Experiments. Volume 2: Advanced Experimental Design By Klaus Hinkelmann and Oscar Kempthorne ISBN 0-471-55177-5 Copyright  2005 John Wiley & Sons, Inc.

728

729

APPENDIX D

2. 4 × 2 × 2, k = 8 (Li, 1944)

Replication 1 Block

Replication 2 Block

Replication 3 Block

1

2

1

2

1

2

000 011 100 111 201 210 301 310

001 010 101 110 200 211 300 311

000 011 101 110 201 210 300 311

001 010 100 111 200 211 301 310

000 011 101 110 200 211 301 310

001 010 100 111 201 210 300 311

λ(000) = 3

h(000) =

2

λ(100) = 1

h(100) =

λ(010) = 0

h(010) =

λ(110) = 2

h(110) =

λ(001) = 0

h(001) =

λ(101) = 2

h(101) =

λ(011) = 3

h(011) =

λ(111) = 1

h(111) =

1 4 1 2 − 18 1 2 − 18 − 14 1 8

E(111) = 23 , all other E(z) = 1, E =

14 15 .

3. 4 × 4, k = 4 (Yates, 1937b)

Replication 1 Block

Replication 2 Block

Replication 3 Block

1

2

3

4

1

2

3

4

1

2

3

4

00 11 22 33

01 10 23 32

02 13 20 31

03 12 21 30

00 12 23 31

01 13 22 30

02 10 21 33

03 11 20 32

00 13 21 32

01 12 20 33

02 11 23 30

03 10 22 31

730

APPENDIX D

λ(00) = 3 λ(10) = 0 λ(01) = 0 λ(11) = 1

h(00) = 2 h(10) = 14 h(01) = 14 h(11) = − 14

E(10) = E(01) = 1, E(11) = 23 , E = 45 . 4. 4 × 3 × 2, k = 12 (Li, 1944) Replication 1 Block

Replication 2 Block

Replication 3 Block

1

2

1

2

1

2

000 011 021 101 110 120 200 211 221 301 310 320

001 010 020 100 111 121 201 210 220 300 311 321

001 010 021 100 111 120 201 210 221 300 311 320

000 011 020 101 110 121 200 211 220 301 310 321

001 011 020 100 110 121 201 211 220 300 310 321

000 010 021 101 111 120 200 210 221 301 311 320

Replication 4 Block

Replication 5 Block

Replication 6 Block

1

2

1

2

1

2

001 010 020 100 111 121 200 211 221 301 310 320

000 011 021 101 110 120 201 210 220 300 311 321

000 011 020 101 110 121 201 210 221 300 311 320

001 010 021 100 111 120 200 211 220 301 310 321

000 010 021 101 111 120 201 211 220 300 310 321

001 011 020 100 110 121 200 210 221 301 311 320

731

APPENDIX D

Replication 7 Block

Replication 8 Block

Replication 9 Block

1

2

1

2

1

2

000 011 021 100 111 121 201 210 220 301 310 320

001 010 020 101 110 120 200 211 221 300 311 321

001 010 021 101 110 121 200 211 220 300 311 320

000 011 020 100 111 120 201 210 221 301 310 321

001 011 020 101 111 120 200 210 221 300 310 321

000 010 021 100 110 121 201 211 220 301 311 320

E(101) =

λ(000) = 9

h(000) =

λ(100) = 3

h(100) =

λ(010) = 3

h(010) =

λ(110) = 5

h(110) =

λ(001) = 0

h(001) =

λ(101) = 6

h(101) =

λ(011) = 6

h(011) =

λ(111) = 4

h(111) =

26 27 , E(111)

=

23 27 ,

23 3 1 3 1 3 1 − 12 2 3 − 16 − 16 − 13

all other E(z) = 1, E =

22 23 .

Using replicates 1, 2, 3 only will yield a design with the same relative efficiencies, but the design is not a BFD. 5. 4 × 4 × 3, k = 12 (Li, 1944) Replication 1 Block

Replication 2 Block

Replication 3 Block

1

2

3

4

1

2

3

4

1

2

3

4

000 110 220

010 100 230

020 130 200

030 120 210

000 130 210

030 100 220

010 120 200

020 110 230

000 120 230

020 100 210

030 110 200

010 130 220

732

APPENDIX D

Replication 1 Block

Replication 2 Block

Replication 3 Block

1

2

3

4

1

2

3

4

1

2

3

4

330 001 111 221 331 002 112 222 332

320 011 101 231 321 012 102 232 322

310 021 131 201 311 022 132 202 312

300 031 121 211 301 032 122 212 302

320 001 131 211 321 002 132 212 322

310 031 101 221 311 032 102 222 312

330 011 121 201 331 012 122 202 332

300 021 111 231 301 022 112 232 302

310 001 121 231 311 002 122 232 312

330 021 101 211 331 022 102 212 332

320 031 111 201 321 032 112 202 322

300 011 131 221 301 012 132 222 302

λ(000) = 3

h(000) =

3

λ(100) = 0

h(100) =

0

λ(010) = 0

h(010) =

0

λ(110) = 1

h(110) =

0

λ(001) = 3

h(001) = − 13

λ(101) = 0

h(101) =

λ(011) = 0

h(011) =

λ(111) = 1

h(111) =

1 12 1 12 1 − 12

E(110) = 23 , all other E(z) = 1, E =

44 47 .

6. 3 × 3 × 2, k = 6 (Yates, 1937b) Replication 1 Block

Replication 2 Block

Replication 3 Block

Replication 4 Block

1

2

3

1

2

3

1

2

3

1

2

3

100 010 220 201 021 111

200 020 110 001 121 211

000 120 210 101 011 221

200 020 110 101 011 221

000 120 210 201 021 111

100 010 220 001 121 211

100 210 020 201 011 121

200 010 120 001 111 221

000 110 220 101 211 021

200 010 120 101 211 021

000 110 220 201 011 121

100 210 020 001 111 221

733

APPENDIX D

λ(000) = 4

h(000) =

λ(100) = 0

h(100) =

λ(010) = 0

h(010) =

λ(110) = 2

h(110) =

λ(001) = 0

h(001) =

λ(101) = 2

h(101) =

λ(011) = 2

h(011) =

λ(111) = 1

h(111) =

5 2 1 2 1 2 − 16 1 2 − 16 − 16 − 16

E(110) = 78 , E(111) = 58 , all other E(z) = 1, E =

15 17 .

If only replications 1 and 2 or 3 and 4 are used, then the relative information for A1 × A2 is 34 and A1 × A2 × A3 is 14 , but the design is not a BFD. 7. 3 × 3 × 2 × 2, k = 12 (Li, 1944) Replication 1 Block

Replication 2 Block

1

2

3

1

2

3

0100 0111 0201 0210 1001 1010 1200 1211 2000 2011 2101 2110

0000 0011 0101 0110 1100 1111 1201 1210 2001 2010 2200 2211

0001 0010 0200 0211 1000 1011 1101 1110 2100 2111 2201 2210

0101 0110 0200 0211 1000 1011 1201 1210 2001 2010 2100 2111

0001 0010 0100 0111 1101 1110 1200 1211 2000 2011 2201 2210

0000 0011 0201 0210 1001 1010 1100 1111 2101 2110 2200 2211

Replication 3 Block

Replication 4 Block

1

2

3

1

2

3

0101 0110 0200

0000 0011 0201

0001 0010 0100

0100 0111 0201

0001 0010 0200

0000 0011 0101

734

APPENDIX D

Replication 3 Block

Replication 4 Block

1

2

3

1

2

3

0211 1001 1010 1100 1111 2000 2011 2201 2210

0210 1101 1110 1200 1211 2001 2010 2100 2111

0111 1000 1011 1201 1210 2101 2110 2200 2211

0210 1000 1011 1101 1110 2001 2010 2200 2211

0211 1100 1111 1201 1210 2000 2011 2101 2110

0110 1001 1010 1200 1211 2100 2111 2201 2210

λ(0000) = 4

h(0000) =

λ(1000) = 0

h(1000) =

λ(0100) = 0

h(0100) =

λ(1100) = 2

h(1100) =

λ(0010) = 0

h(0010) =

λ(1010) = 2

h(1010) =

λ(0110) = 2

h(0110) =

λ(1110) = 1

h(1110) =

λ(0001) = 0

h(0001) =

λ(1001) = 2

h(1001) =

λ(0101) = 2

h(0101) =

λ(1101) = 1

h(1101) =

λ(0011) = 4

h(0011) =

λ(1011) = 0

h(1011) =

λ(0111) = 0

h(0111) =

λ(1111) = 2

h(1111) =

5 2 1 2 1 2 − 16 3 4 − 14 − 14 1 12 3 4 − 14 − 14 1 12 − 12 1 6 1 6 − 16

E(1100) = 78 , E(1111) = 58 , all other E(z) = 1. If only replications 1 and 2 or 3 and 4 are used, the relative information for A1 × A2 remains 78 and for A1 × A2 × A3 × A4 remains 58 , but these designs are not BFDs.

735

APPENDIX D

8. 5 × 2 × 2, k = 10 (Shah, 1960)

1

2

3

4

000 011 100 111 210 201 310 301 410 401

010 001 100 111 200 211 310 301 410 401

010 001 110 101 200 211 300 311 410 401

010 001 110 101 210 201 300 311 400 411

E(011) =

Block 5 6 000 011 110 101 210 201 310 301 400 411

010 001 100 111 210 201 300 311 400 411

7

8

9

10

000 011 110 101 200 211 310 301 400 411

000 011 100 111 210 201 300 311 410 401

010 001 100 111 200 211 310 301 400 411

000 011 110 101 200 211 300 311 410 401

λ(000) = 5

h(000) =

λ(100) = 2

h(100) =

λ(010) = 0

h(010) =

λ(110) = 3

h(110) =

λ(001) = 0

h(001) =

λ(101) = 3

h(101) =

λ(011) = 5

h(011) =

λ(111) = 2

h(111) =

24 25 , E(111)

=

19 25 ,

38 10 2 10 6 10 1 − 10 6 10 1 − 10 3 − 10 2 − 10

all other E(z) = 1, E =

18 19 .

APPENDIX E

Exercises

CHAPTER 1 1.1 Consider the following design with t = 7 treatments in blocks of size 3 and 4: Block

Treatments

1

2

3

4

5

6

7

2

3

3

5

1

7

3

4

5

6

4

6

2

1

1

2

4

7

5

6

7

4

1

2 (a) (b) (c) (d)

3

Give the parameters for this design. Write out the incidence matrix. Write out the C matrix of (1.9). Show that all differences τi − τi are estimable.

1.2 Show that CI = O for C of (1.9). 1.3 The experiment of Exercise 1.1 was originally planned for blocks of size 3, but it turned out that some of the blocks had actually 4 experimental units. The investigator decided to replicate one of the treatments assigned to those blocks. (a) Sketch the ANOVA table for this experiment, giving sources of variation and d.f. Design and Analysis of Experiments. Volume 2: Advanced Experimental Design By Klaus Hinkelmann and Oscar Kempthorne ISBN 0-471-55177-5 Copyright  2005 John Wiley & Sons, Inc.

736

737

APPENDIX E

(b) Discuss how the error d.f. can be divided into two sets and what the advantage of that would be. 1.4 Consider the following incomplete block design with t = 6, b = 6, and k = 2:

Treatments

1

2

Block 3 4

1 2

3 1

2 3

4 6

5

6

5 6

5 4

(a) Write out the C matrix. (b) Show that this is a disconnected design. (c) Identify a set of estimable functions of the treatment effects. 1.5 Consider the following data obtained from an experiment using an incomplete block design with t = 9, b = 6, k = 3, and 2 subsamples (the numbers in parentheses indicate the treatments): 1 2 3 Block

4 5 6

53.5 , 54.8 (6) 53.1 , 52.7 (3) 57.2 , 56.5 (9) 53.7 , 52.9 (4) 54.5 , 53.3 (3) 48.9 , 47.8 (8)

53.2 , 54.0 (4) 58.6 , 57.1 (1) 55.0 , 55.9 (7) 53.6 , 54.6 (2) 52.8 , 53.2 (7) 53.5 , 54.9 (6)

57.7 , 56.4 (7) 53.9 , 55.1 (2) 51.5 , 53.2 (8) 57.9 , 56.8 (9) 53.3 , 55.0 (5) 56.7 , 55.4 (1)

(a) Write out a model for this data set. (b) Obtain the intrablock analysis using SAS. (c) Obtain the standard errors for all treatment comparisons τi − τi . 1.6 Consider an incomplete block design with t treatments, b blocks of size k, r replications per treatment, and m subsamples per experimental unit. (a) Explain how the derivations in Sections 1.8.2 and 1.8.3 need to be modified for this case. (b) Using the data from Exercise 1.5 obtain the combined analysis, including treatment comparisons. (c) Compare the results with those of Exercise 1.5.

738

APPENDIX E

CHAPTER 2 2.1 Consider the BIB design Block 1 2 3 4 5 6

1 2 1 1 1 4

4 6 3 2 5 5

6 8 8 3 7 6

7 9 9 4 8 9

Block 7 8 9 10 11 12

2 2 3 1 2 3

3 4 5 2 3 4

6 5 7 5 5 7

7 8 9 7 6 9

Block 13 14 15 16 17 18

1 1 1 4 3 2

2 5 3 6 4 7

4 6 6 7 5 8

9 9 8 8 8 9

(a) Give the parameters of this design. (b) Write out its C matrix. 2.2 (a) Describe a scenario where you may want to use the design given in Exercise 2.1 but in such a form that each treatment occurs exactly 4 times in the first 2 positions and 4 times in the last 2 positions over the 18 blocks. (b) Write out the Latinized form of this design. (c) Give the SAS statements for analyzing data using the design in (b). 2.3 (a) Show that the design given in Exercise 2.1 is a 4-resolvable BIB design. (b) Give the SAS statements for analyzing data using the design in (a). 2.4 (a) Show that the design obtained in Exercise 2.3 can be Latinized. (b) Give the SAS statements for analyzing data using the design obtained in (i). 2.5 Investigate whether the design given in Exercise 2.1 is locally resistant w.r.t. any of the treatments. CHAPTER 3 3.1 Give a proof of Theorem 3.2. 3.2 Obtain the complement of the BIB design given in Exercise 2.1 and verify that it is a BIB design. 3.3 (a) Using the successive diagonalizing method construct a BIB for t = 25 treatment in blocks of size 5. (b) Obtain the C matrix for this design. (c) Suppose that the last replicate is missing from the design. Investigate the properties of the residual design w.r.t. estimating treatment differences τi − τi .

APPENDIX E

739

3.4 Construct the BIB design #25 in Table 3.1. 3.5 Construct the BIB design #18 in Table 3.1. 3.6 Construct the BIB design #22 in Table 3.1. CHAPTER 4 4.1 Consider the rectangular PBIB design of Example 4.8. (a) Write out explicitly Eqs. (4.34) and obtain a solution of these equations. (b) Give explicit expressions for the variances of simple treatment comparisons. (c) Generate data for this design and, using the g inverse obtained from SAS PROC GLM, verify the results obtained in (b). 4.2 Write out the association scheme for the L4 -PBIB(2) design for t = 9 treatments. 4.3 (a) Write out the association scheme for the EGD/(7)-PBIB design with t = 18 treatments. (b) Give the parameters of this design. (c) Give the P matrices for this design. CHAPTER 5 5.1 Verify that the design D of Example 5.1 is a PBIB(2) design and give its parameters, including the P matrices. 5.2 Using the PG(3, 2) construct a GD-PBIB(2) with t = 2 · 6 = 12 treatments in 8 blocks of size 6. 5.3 Show that the dual of the BIB design #8 in Table 3.1 is a triangular PBIB(2) design. 5.4 Construct a L2 -PBIB(2) design for t = 25 treatments in 20 blocks of size 5. 5.5 (a) Construct a cyclic PBIB design for t = 7 treatments in b = 7 blocks of size k = 4. (b) Obtain the efficiency factor of this design. (c) Check whether there exists a more efficient cyclic PBIB design than the design obtained in (a). 5.6 Construct a balanced array BA[15, 6, 3, 2] from OA[18, 7, 3, 2; 2] and give its index set. 5.7 Using the method of balanced arrays construct an EGD/3-PBIB design for t = 12 treatments in 12 blocks of size 3.

740

APPENDIX E

5.8 Using the method of balanced arrays construct an EGD/7-PBIB design for t = 12 treatments in 12 blocks of size 2. CHAPTER 6 6.1 (a) Construct an α design for t = 15 treatments in r = 3 replicates and blocks of size k = 5. (b) Describe the properties of this design. (c) Obtain the efficiency factor of this design. 6.2 Construct an α design for t = 19 treatments in blocks of size 5 and 4, with r = 2 replicates. 6.3 Construct a generalized cyclic incomplete block design for t = 15 treatments in blocks of size 5. 6.4 (a) Show that the design of Example 6.6 with r = 3 replicates is a (0, 1) design. (b) Obtain the efficiency factor of this design. 6.5 Consider the BTIB design with t = 4 test treatments and blocks of size 3 with generator designs   0 0 0 0 0 0   GD1 =  1 1 1 2 2 3  

2 3

4 3 4

4

0 0

1 1 1

1 2 2

 GD2 =  1 3 



2 2 3

 3 3 4 

2 4

3 4 4

4 3 4

1 1

1 2

3 4

4 4

 GD3 =  2 2



 3 3 

(a) Using the generator designs given above, construct two BTIB designs with b = 12 blocks. (b) Give the parameters of both designs and compare their efficiency factors. 6.6 Consider the GD-PBIB(2) design with 6 treatments and 9 blocks of size 2: 1

3 5 1 6

5 1 3

5

2

4 6 3 2

4 4 6

2

741

APPENDIX E

(a) By inspection of the design above give its parameters and the association structure. (b) Using the PBIB design above, obtain a PBTIB design for t = 5 test treatments. (c) Give the parameters of the design obtained in (b). 6.7 (a) Using the PBIB design of Exercise 6.6, obtain a PTBIB design for t = 4 test treatments. (b) Give the parameters of this design. 6.8 Consider the GD-PBIB(2) design for t = 6 treatments in b = 6 blocks of size 4: 1 4 2

5

4

1 5 2

2 5 3

6

5

2 6 3

3 6 1

4

6

3 4 1

(a) Investigate the properties of this design as a row–column design. (b) Obtain the efficiency factor of the design obtained in (a). CHAPTER 7 7.1 For the 24 factorial give explicit expressions for the true response of the treatment combination x =(0, 1, 0, 1) using (7.42) and (7.49). 7.2 For the 25 factorial write out explicitly the contrast that defines the 4-factor interaction ABCD. 7.3 For the 23 factorial with r replications per treatment and m subsamples per experimental unit write out: (a) An expression for the interaction AB. (b) An expression for the sum of squares for AB. (c) The ANOVA table, including sources of variation, d.f., SS, E(MS). 7.4 Consider the 26 factorial with one replication per treatment combination. Assume that all interactions involving three or more factors are negligible. (a) Write out the ANOVA table including sources of variation, d.f., SS, E(MS). (b) Assume that there are m = 2 subsamples per treatment combination. Write out the ANOVA table, including sources of variation, d.f., SS, E(MS). (c) For the situation described in (b) give the input statements for SAS PROC GLM and SAS PROC MIXED to perform both the intrablock and combined analyses.

742

APPENDIX E

CHAPTER 8 8.1 (a) Determine the intrablock subgroup for the 25 factorial in blocks of size 8, confounding interactions involving three or more factors. (b) Write out the remaining blocks. 8.2 Generate the design of Exercise 8.1 using SAS PROC FACTEX. 8.3 Consider the design generated in Exercise 8.1 with r = 2 replications. (a) Assuming that all interactions involving three or more factors are negligible, write out the ANOVA table, including sources of variation, d.f., SS, and E(MS). (b) Suppose that for each experimental unit we have m = 2 subsamples. Modify the ANOVA table in (a) appropriately. 8.4 For the situations described in Exercise 8.3 give the SAS statements for performing the respective analyses, using SAS PROC GLM and SAS PROC MIXED.

CHAPTER 9 9.1 Consider the 25 factorial with 3 replicates and in blocks of size 4. Without confounding main effects construct a design that provides as much and as uniform as possible information about 2-factor interactions. 9.2 For the design obtained in Exercise 9.1, assuming that 4- and 5-factor interactions are negligible, give the ANOVA table, including sources of variation, d.f., SS, and E(MS). 9.3 For the design obtained in Exercise 9.1 give expressions for the combined estimators of all 2-factor interactions and the variances of these estimators. 9.4 Discuss some possibilities of arranging a 26 factorial in two 8 × 8 squares, giving the actual designs and the relative information provided for main effects and 2-factor interactions.

CHAPTER 10 10.1 For the 34 factorial write out all main effects and interactions together with their 2-d.f. components. 10.2 For the 34 factorial write out the parametrization (10.15) for the true response of the treatment combination x = (0, 1, 1, 2). 10.3 Construct a design for the 34 factorial in blocks of size 9 without confounding main effects and 2-factor interactions.

743

APPENDIX E

10.4 Consider the 34 factorial in blocks of size 3. (a) Suppose two replicates are available. Using partial confounding, construct a design that obtains at least partial information on all main effects and 2-factor interaction components. (b) Give the relative information for all main effects and 2-factor interaction components. (c) Assuming that all 3- and 4-factor interactions are negligible, sketch the ANOVA table for the design obtained in (a), giving sources of variation, d.f., SS, and E(MS). (d) Give the SAS statements for performing the analysis of an experiment using the design obtained in (a) for PROC GLM (intrablock analysis) and PROC MIXED (combined analysis).

CHAPTER 11 11.1 Consider the 52 factorial in 10 blocks of size 5. (a) Obtain a plan that yields full information on main effects and at least partial information on all interaction components. (b) Sketch the ANOVA table, including sources of variation, d.f., and SS. (c) Give the parametrization for a(1 1 1 1 1) − a(0 0 0 0 0) and give the variance of the estimator for this difference in yields. 11.2 As an alternative to the design obtained in Exercise 11.1 consider the design based on the PBIB-L2 design LS 51 (Clatworthy, 1973) given by Block 1 2 3 4 5

Block 1 6 11 16 21

2 3 7 8 12 13 17 18 22 23

4 9 14 19 24

5 10 15 20 25

6 7 8 9 10

1 2 3 4 5

6 7 8 9 10

11 12 13 14 15

16 17 18 19 20

21 22 23 24 25

(a) Set up a correspondence between the treatments 1, 2, . . ., 25 and the factorial treatments (x1 , x2 , x3 , x4 , x5 ). (b) Explain the major difference between the two designs with respect to the estimation of main effects and interaction effects. (c) Explain how you would compare the overall efficiency of the two designs. 11.3 Consider the 43 factorial in blocks of size 16. (a) Assuming that 3-factor interactions are negligible, obtain a plan that yields full information on main effects and 2-factor interactions.

744

APPENDIX E

(b) Suppose that 8 blocks are available for the experiment. Write out the ANOVA table, including sources of variation, d.f., and SS. 11.4 (a) Apply Theorem 11.7 to the case of the 4n factorial and write out the Fisher plan for the maximum number of factors. (b) For the plan obtained in (a) identify the resulting system of confounding. 11.5 (a) For the 43 factorial in blocks of size 8 obtain a system of confounding with orthogonal factorial structure. (b) Obtain the efficiency factor of the design.

CHAPTER 12 12.1 Consider the 22 × 33 factorial in 18 blocks of size 6. (a) Assuming that interactions involving three or more factors are negligible, obtain a design using the Kronecker product method. (b) Explain why the design obtained in (a) is or is not satisfactory, and if it is not satisfactory what you might want to do to fix the problem. 12.2 Consider the 22 × 33 factorial in blocks of size 6. (a) Using the GC/n method construct a system of confounding. (b) Obtain the loss of d.f. for the design obtained in (a). (c) Compare the design obtained in (a) with that obtained in Exercise 12.1. 12.3 (a) Apply the method of finite rings to the 22 × 33 factorial in blocks of size 6. (b) Obtain the variance for the comparison between the treatments (1, 1, 2, 2, 2) and (0, 0, 0, 0, 0). 12.4 Consider the 22 × 42 factorial in blocks of size 8. (a) Construct a balanced factorial design such that all treatment contrasts are estimable. (b) For the design of (a) obtain the efficiency factor.

CHAPTER 13 13.1 Consider a 28−2 design. (a) Give an identity relationship such that all main effects are strongly clear and all 2-factor interactions are either clear or strongly clear. (b) Obtain the design based on the identity relationship in (a). (c) Give the resolution of the design.

APPENDIX E

745

13.2 Consider the design in Exercise 13.1. Suppose that for the actual experiment blocks of size 16 need to be used. (a) Identify a suitable system of confounding. (b) Obtain the actual design. (c) For the design in (b) sketch the ANOVA table, including sources of variation and d.f., assuming that interactions involving four or more factors are negligible. 13.3 Consider a 36−2 IV factorial based on the identity relationship I = ABCD = AB 2 EF = all GI. (a) Complete the identity relationship. (b) Write out the alias structure for all main effects and 2-factor interaction components. (c) Identify all the clear effects. (d) Obtain the actual design. 13.4 For the design of Exercise 13.3 suppose that blocks of size 9 need to be used. (a) Obtain a suitable system of confounding and state any assumptions you are making. (b) For the arrangement in (a) sketch the ANOVA, giving sources of variation and d.f., assuming that all interactions involving three or more factors are negligible. 13.5 Consider the 26−2 IV fraction in 4 blocks. (a) Using the independent defining words given in Table 13.12, obtain the actual design. (b) Suppose that the plan in (a) is replicated three times. Sketch the ANOVA, including sources of variation and d.f. (c) Rather than replicating the plan in (a) three times, discuss other possibilities of carrying out the experiment.

CHAPTER 14 14.1 Construct a saturated orthogonal main effect plan for the 27 factorial. 14.2 Construct an OMEP for the 34 factorial in 9 runs. 14.3 (a) Construct an OMEP for the 24 factorial in 9 runs. (b) Identify the alias structure for this design. 14.4 Using the OMEPs obtained in Exercises 14.1 and 14.2, construct an OMEP for the 27 × 34 factorial. 14.5 (a) Construct an OMEP for the 42 × 3 × 22 factorial.

746

APPENDIX E

(b) Suppose the four- and three-level factors are quantitative factors with equidistant levels. Show that the estimators for the linear effects of a four-level factor and the three-level factor are uncorrelated. 14.6 (a) Construct an OMEP for the 4 × 32 × 22 factorial. (b) Suppose the plan in (a) is replicated twice. Sketch the ANOVA table, including sources of variation and d.f. 14.7 Consider the following main effect plan for the 24 factorial in five runs: Run

A

B

C

D

1 2 3 4 5

0 0 1 1 1

0 1 0 1 1

0 1 1 0 1

0 1 1 1 0

(a) Is this an OMEP? (b) Show how you would estimate the main effects. (c) Obtain the efficiency factor of this design.

CHAPTER 15 15.1 Construct a supersaturated design for 14 factors in 8 runs. 15.2 (a) Construct a supersaturated design for 20 factors in 12 runs. (b) Obtain E(s 2 ) for this design and compare it to the lower bound (15.9). 15.3 (a) Construct a three-level supersaturated design for 12 factors in 12 runs. (b) Obtain the av χ 2 and max χ 2 values of (15.3) and (15.4), respectively.

CHAPTER 16 16.1 (a) Using the results of Table 16.1, write out the resolution III.1 design for n = 5 factors. (b) Verify that the conditions of Theorem 16.5 are satisfied. (c) Show explicitly that, individually, each 2-factor interaction is estimable. 16.2 (a) Obtain the resolution V.1 design for n = 4 factors. (b) Verify that the conditions of Theorem 16.4 is satisfied. (c) Show explicitly that, individually, each 3-factor interaction is estimable.

747

APPENDIX E

16.3 Find the search probabilities for the design obtained in Exercise 16.1. 16.4 Fine the search probabilities for the design obtained in Exercise 16.2.

CHAPTER 17 17.1 Consider the case of three design factors A1 , A2 , A3 , each at three levels, and two noise factors B1 , B2 , also each at three levels. (a) Construct the smallest cross array design such that all main effects are estimable. (b) Suppose that all five factors are quantitative factors and that only linear × linear interactions between Ai and Bj (i = 1, 2, 3; j = 1, 2) are important. Sketch the ANOVA table for the design in (a), including sources of variation and d.f. 17.2 Find an alternative single array design to the design obtained in Exercise 17.1 that uses fewer runs. 17.3 In certain cases when the design and noise factors have two levels it is possible to utilize a 2n− fractional factorial to construct a single array design. For example, for three design factors A1 , A2 , A3 and three noise factors B1 , B2 , B3 the 26−2 fraction based on I = A1 A2 A3 B2 = A1 A2 A3 B1 B3 = B1 B2 B3 yields a suitable design. (a) Write out the design in terms of control and noise factor level combinations. (b) Identify the design–noise factor interactions that are clear. 17.4 (a) Use the method described in Exercise 17.3 to obtain a single array for four design factors and two noise factors, each at two levels. (b) Identify all the clear effects. (c) What additional assumptions will you have to make to estimate all main effects of the control factors and all control × noise factor interactions?

CHAPTER 18 18.1 (a) Construct a balanced square lattice for 25 treatments in blocks of size 5. (b) Sketch the ANOVA table for this design, including sources of variation and d.f. (c) Show that this design is a BIB design.

748

APPENDIX E

18.2 (a) Construct a triple three-dimensional lattice for 64 treatments in blocks of size 4. (b) Sketch the ANOVA table for this design, including sources of variation and d.f. (c) Consider the B | T-ANOVA and indicate how you would estimate the weights w and w . 18.3 Construct a simple square lattice for 81 treatments in blocks of size 9. 18.4 (a) Construct a balanced lattice square for 16 treatments. (b) Sketch the ANOVA table for this design, including sources of variation and d.f. 18.5 Construct a triple rectangular lattice for 30 treatments in blocks of size 5.

CHAPTER 19 19.1 Construct a balanced CO(5, 10, 3). 19.2 Consider the irreducible BIB design for 4 treatments in blocks of size 3. Use this design to construct the CO(4, 24, 3). 19.3 (a) Consider the PBIB(2) design in (5.9) and use it as the generator design to construct the CO(6, 12, 3). (b) Obtain the variances for all simple treatment comparisons. 19.4 Construct the strongly balanced CO(4, 8, 3). 19.5 Suppose we have a 24 factorial treatment structure. Assume that the 4-factor interaction is negligible. (a) Construct the CO(24 , 24 , 23 ). (b) Explain how you would estimate the direct main effects and 2-factor interactions.

References

Abel, R. J. R., A. E. Brouwer, C. J. Colbourn and J. H. Dinitz (1996). Mutually orthogonal latin squares (MOLS). In: The CRC Handbook of Combinatorial Designs. C. J. Colbourn and J. H. Dinitz (eds.). Boca Raton, FL : CRC, pp. 111–142. Abraham, B., H. Chipman and K. Vijayan (1999). Some risks in the construction and analysis of supersaturated designs. Technometrics, 41, 135–141. Addelman, S. (1961). Irregular fractions of the 2n factorial experiment. Technometrics, 3, 479–496. Addelman, S. (1962a). Orthogonal main-effect plans for asymmetrical factorial experiments. Technometrics, 4, 21–46 (Erratum 4, 440). Addelman, S. (1962b). Symmetrical and asymmetrical fractional factorial plans. Technometrics, 4, 47–58. Addelman, S. and O. Kempthorne (1961). Some main-effect plans and orthogonal arrays of strength two. Ann. Math. Statist., 32, 1167–1176. Adhikary, B. (1966). Some types of m-associate PBIB association schemes. Calcutta Statist. Assoc. Bull., 15, 47–74. Adhikary, B. (1967). A new type of higher associate cyclical association schemes. Calcutta Statist. Assoc. Bull., 16, 40–44. Afsarinejad, K. (1983). Balanced repeated measurement designs. Biometrika, 70, 199–204. Afsarinejad, K. (1990). Repeated measurement designs—A review. Commun. Statist.— Theory Meth., 19, 3985–4028. Afsarinejad, K. and A. S. Hedayet (2002). Repeated measurements designs for a model with self and simple mixed carryover effects. J. Statist. Plann. Inf., 106, 449–459. Aggarwal, K. R. (1974). Some higher-class PBIB designs and their applications as confounded factorial experiments. Ann. Inst. Statist. Math., 26, 315–323. Aggarwal, K. R. and T. P. Singh (1981). Methods of construction of balanced arrays with application to factorial designs. Calcutta Statist. Assoc. Bull., 30, 89–93. Atkinson, A. C. and A. N. Donev (1992). Optimum Experimental Design. Oxford: Clarendon. Design and Analysis of Experiments. Volume 2: Advanced Experimental Design By Klaus Hinkelmann and Oscar Kempthorne ISBN 0-471-55177-5 Copyright  2005 John Wiley & Sons, Inc.

749

750

REFERENCES

Baksalary, J. K. and P. D. Puri (1990). Pairwise-balanced, variance-balanced and resistant incomplete block designs revisited. Ann. Inst. Statist. Math., 42, 163–171. Balaam, L. N. (1968). A two-period design with t 2 experimental units. Biometrics, 24, 61–73. Banerjee, K. S. (1970). Some observations on the White-Hultquist procedure for the construction of confounding plans for mixed factorial designs. Ann. Math. Statist., 41, 1108–1113. Banerjee, K. S. (1975). Weighing Designs. New York: Dekker. Baumert, L. D., S. W. Golomb and M. Hall, Jr. (1962). Discovery of an Hadamard matrix of order 92. Bull. Am. Math. Soc., 68, 237–238. Bechhofer, R. E. and A. C. Tamhane (1981). Incomplete block designs for comparing treatments with a control: General theory. Technometrics, 23, 45–57 (Corrigendum 24, 171). Bechhofer, R. E. and A. C. Tamhane (1983a). Incomplete block designs for comparing treatments with a control (II): Optimal designs for p = 2(1)6, k = 2 and p = 3, k = 3. Sankhya, Ser. B, 45, 193–224. Bechhofer, R. E. and A. C. Tamhane (1983b). Design of experiments for comparing treatments with a control: Tables of optimal allocations of observations. Technometrics, 25, 87–95. Bechhofer, R. E. and A. C. Tamhane (1985). Tables of admissible and optimal balanced treatment incomplete block (BTIB) designs for comparing treatments with a control. Selected Tables Math. Statist., 8, 41–139. Berenblut, I. I. (1964). Change-over design with complete balance for first residual effects. Biometrics, 20, 707–712. Berenblut, I. I. (1967). A change-over design for testing a treatment factor at four equally spaced levels. J. Roy. Statist. Soc., B, 29, 370–373. Berenblut, I. I. (1968). Change-over designs balanced for the linear component of first residual effects. Biometrika, 55, 297–303. Beyer, W. H. (ed.) (1991). CRC Standard Probability and Statistics Tables and Formulae. Boca Raton, FL : CRC. Bhattacharya, C. G. (1998). Goodness of Yates-Rao procedure for recovery of inter-block information. Sankhya A, 60, 134–144. Bisgaard, S. (1994). A note on the definition of resolution for blocked 2k−p designs. Technometrics, 36, 308–311. Blaisdell, E. A. and D. Raghavarao (1980). Partially balanced changeover designs based on m-associate PBIB designs. J. Roy. Statist. Soc., B, 42, 334–338. Booth, K. H. V. and D. R. Cox (1962). Some systematic supersaturated designs. Technometrics, 4, 489–495. Bose, R. C. (1939). On the construction of balanced incomplete block designs. Ann. Eugenics, 9, 353–399. Bose, R. C. (1942). A note on the resolvability of balanced incomplete block designs. Sankhya, 6, 105–110. Bose, R. C. (1947a). On a resolvable series of balanced incomplete block designs. Sankhya, 8, 249–256. Bose, R. C. (1947b). Mathematical theory of the symmetrical factorial design. Sankhya, 8, 107–166.

REFERENCES

751

Bose, R. C. and K. A. Bush (1952). Orthogonal arrays of strength two and three. Ann. Math. Statist., 23, 508–524. Bose, R. C. and W. S. Connor (1952). Combinatorial properties of group divisible incomplete block designs. Ann. Math. Statist., 23, 367–383. Bose, R. C. and K. Kishen (1940). On the problem of confounding in the general symmetrical factorial design. Sankhya, 5, 21–36. Bose, R. C. and D. M. Mesner (1959). On linear associative algebras corresponding to association schemes of partially balanced designs. Ann. Math. Statist., 30, 21–38. Bose, R. C. and K. R. Nair (1939). Partially balanced incomplete block designs. Sankhya, 4, 337–372. Bose, R. C. and T. Shimamoto (1952). Classification and analysis of partially balanced incomplete block designs with two associate classes. J. Am. Statist. Assoc., 47, 151–184. Bose, R. C., W. H. Clatworthy and S. S. Shrikhande (1954). Tables of partially balanced incomplete block designs with two associate classes. North Carolina Agricultural Experiment Station Technical Bulletin No. 107, Raleigh, NC. Bose, R. C., S. S. Shrikhande and K. N. Bhattacharya (1953). On the construction of group divisible incomplete block designs. Ann. Math. Statist., 24, 167–195. Box, G. E. P. (1988). Signal-to-noise ratios, performance criteria, and transformation. Technometrics, 30, 1–17. Box, G. E. P. and J. S. Hunter (1961a). The 2k−p fractional factorial designs. Part I. Technometrics, 3, 311–351 (Errata 5, 417). Box, G. E. P. and J. S. Hunter (1961b). The 2k−p fractional factorial designs. Part II. Technometrics, 3, 449–458. Box, G. E. P. and R. D. Meyer (1986a). An analysis of unreplicated fractional factorials. Technometrics, 28, 11–18. Box, G. E. P. and R. D. Meyer (1986b). Dispersion effects from fractional designs. Technometrics, 28, 19–27. Box, G. E. P. and K. B. Wilson (1951). On the experimental attainment of optimum conditions. J. Roy. Statist. Soc., B, 13, 1–45. Bradley, J. V. (1958). Complete counterbalancing of immediate sequential effects in a Latin square design. J. Am. Statist. Assoc., 53, 525–528. Bulutoglu, D. A. and C.-S. Cheng (2003). Hidden projection properties of some nonregular fractional factorial designs and their applications. Ann. Statist., 31, 1012–1026. Calinski, T. and S. Kageyama (2000). Block Designs: A Randomization Approach. Volume I: Analysis. New York: Springer. Calvin, L. D. (1954). Doubly balanced incomplete block designs for experiments in which the treatment effects are correlated. Biometrics, 10, 61–88. Carmichael, R. D. (1956). Introduction to the Theory of Groups of Finite Order. New York: Dover. Chakravarti, I. M. (1956). Fractional replication in asymmetrical factorial designs and partially balanced arrays. Sankhya, 17, 143–164. Chakravarti, I. M. (1961). On some methods of construction of partially balanced arrays. Ann. Math. Statist., 32, 1181–1185. Chang, C.-T. (1989). On the construction of hypercubic design. Comm. Statist.—Theory Meth., 18, 3365–3371.

752

REFERENCES

Chang, C.-T. and K. Hinkelmann (1987). A new series of EDG-PBIB designs. J. Statist. Plann. Inf., 16, 1–13. Chang, L.-C. and W.-R. Liu (1964). Incomplete block designs with square parameters for which k ≤ 10 and r ≤ 10. Sci. Sinica, 13, 1493–1495. Chang, L.-C., C.-W. Liu and W.-R. Liu (1965). Incomplete block designs with triangular parameters for which k ≤ 10 and r ≤ 10. Sci. Sinica, 14, 329–338. Chatterjee, K., L.-Y. Deng and D. K. J. Lin (2001). Two-level search design for main-effect plus two plan. Metrika, 54, 233–245. Chatterjee, K. and R. Mukerjee (1986). Some search designs for symmetric and asymmetric factorials. J. Statist. Plann. Inf., 13, 357–363. Chatterjee, K. and R. Mukerjee (1993). Search designs for estimating main effects and searching several two-factor interactions in general factorials. J. Statist. Plann. Inf., 35, 131–138. Chen, H. and C.-S. Cheng (1999). Theory of optimal blocking of 2n−m designs. Ann. Statist., 27, 1948–1973. Chen, H. H. and C.-S. Cheng (2004). Aberration, estimation capacity and estimation index. Statist. Sinica, 14, 203–215. Cheng, C.-S. (1996). Optimal design: Exact theory. In: Handbook of Statistics, Volume 13. S. Ghosh and C. R. Rao (eds.). Amsterdam: Elsevier Science, pp. 977–1006. Cheng, C.-S. (1997). E(s 2 )-optimal supersaturated designs. Statist. Sinica, 7, 929–939. Cheng, C.-S., D. M. Steinberg and D. X. Sun (1999). Minimum aberration and maximum estimation capacity. J. Roy. Statist. Soc., B, 61, 85–93. Cheng, C.-S. and C.-F. Wu (1980). Balanced repeated measurement designs. Ann. Statist., 8, 1272–1283 [Correction 11, (1983), 349]. Chopra, D.V. and J. N. Srivastava (1974). Optimal balanced 28 fractional factorial designs of resolution V, 37 ≤ N ≤ 51. Sankhya, A, 36, 41–52. Chopra, D.V. and J. N. Srivastava (1975). Optimal balanced 27 fractional factorial designs of resolution V, 43 ≤ N ≤ 48. Sankhya, B, 37, 429–447. Clatworthy, W. H. (1955). Partially balanced incomplete block designs with two associate classes and two treatments per block. J. Res. National Bur. Standards, 54. Clatworthy, W. H. (1956). Contributions on partially balanced incomplete block designs with two associate classes. NBS Appl. Math. Series, 47. Clatworthy, W. H. (1967). Some new families of partially balanced designs of the Latin square type and related designs. Technometrics, 9, 229–244. Clatworthy, W. H. (1973). Tables of two-associate-class partially balanced designs. NBS Appl. Math. Series, 63. Cochran, W. G. (1939). Long-term agricultural experiments (with discussion). J. Roy. Statist. Soc., B, 6, 104–148. Cochran, W. G., K. M. Autrey and C. Y. Cannon (1941). A double change-over design for dairy cattle feeding experiments. J. Dairy Sci., 24, 937–951. Cochran, W. G. and G. M. Cox (1957). Experimental Designs, 2nd ed. New York: Wiley. Cochran, W. G. and G. M. Cox (1992). Experimental Designs. New York: Wiley. Cohen, A. and H. B. Sackrowitz (1989). Exact tests that recover interblock information in balanced incomplete block designs. J. Am. Statist. Assoc., 84, 556–559.

REFERENCES

753

Connor, W. S. and W. H. Clatworthy (1954). Some theorems for partially balanced designs. Ann. Math. Statist., 25, 100–112. Constantine, G. (1983). On the trace efficiency for control of reinforced balanced incomplete block designs. J. Roy. Statist. Soc., B, 45, 31–36. Corbeil, R. R. and S. R. Searle (1976). Restricted maximum likelihood (REML) estimation of variance components in the mixed model. Technometrics, 18, 31–38. Cotter, S. C., J. A. John and T. M. F. Smith (1973). Multi-factor experiments in nonorthogonal designs. J. Roy. Statist. Soc., B, 35, 361–367. Cotton, J. W. (1998). Analyzing Within-Subjects Experiments. Mahwah, NJ: Lawrence Earlbaum Associates. Cox, D. R. (1958). The interpretation of the effects of non-additivity in the Latin square. Biometrika, 45, 69–73. Cox D. R. and N. Reid (2000). The Theory of the Design of Experiments. Boca Raton, FL : Chapman and Hall / CRC. Daniel, C. (1959). Use of half-normal plots in interpreting factorial two-level experiments. Technometrics, 1, 311–341. Daniel, C. (1973). One-at-a-time plans. J. Am. Statist. Assoc., 68, 353–360. Das, M. N. (1958). On reinforced incomplete block designs. J. Indian Soc. Agric. Statist., 10, 73–77. David, H. A. (1963). The Method of Paired Comparisons. London: Griffin. David, H. A. (1965). Enumeration of cyclic paired-comparison designs. Am. Math. Monthly, 72, 241–248. David, H. A. and F. W. Wolock (1965). Cyclic designs. Ann. Math. Statist., 36, 1526–1534. Davies, O. L. (ed.) (1956). The Design and Analysis of Industrial Experiments. London: Oliver and Boyd. Davis, A. W. and W. B. Hall (1969). Cyclic change-over designs. Biometrika, 56, 283– 293. Dean, A. M. and J. A. John (1975). Single replication factorial experiments in generalized cyclic designs: II. Asymmetrical arrangements J. Roy. Statist. Soc, B, 37, 72–76. Dénes, J. and A. D. Keedwell (1974). Latin Squares and Their Applications. London: English Universities Press. Dey, A. (1985). Orthogonal Fractional Factorial Design. New York: Halsted. Dunnett, C. W. (1955). A multiple comparison procedure for comparing several treatments with a control. J. Am. Statist. Assoc., 50, 1096–1121. Dunnett, C. W. (1964). New tables for multiple comparisons with a control. Biometrics, 20, 482–491. Eccleston, J. A. and C. A. McGilchrist (1986). Algebra of a row-column design. J. Statist. Plann. Inf., 12, 305–310. Eccleston, J. A. and K. G. Russell (1975). Connectedness and orthogonality in multi-factor designs. Biometrika, 62, 341–345. Eccleston, J. A. and K. G. Russell (1977). Adjusted orthogonality in non-orthogonal designs. Biometrika, 64, 339–345. Ehrenfeld, S. (1953). On the efficiency of experimental design. Ann. Math. Statist., 26, 247–255. Engel, J. (1992). Modelling variation in industrial experiments. Appl. Statist., 41, 579–593.

754

REFERENCES

Fan, J. and R. Li (2001). Variable selection via nonconcave penalized likelihood and its oracle properties. J. Am. Statist. Assoc., 96, 1348–1360. Fang, K.-T., D. K. J. Lin and C.-X. Ma (2000). On the construction of multi-level supersaturated designs. J. Statist. Plann. Inf., 86, 239–252. Federer, W. T. (1955). Experimental Design—Theory and Application. New York: Macmillan. Fichtner, A. (2000). Auswertung eines Mais-Feldexperimentes zur Untersuchung von Genotypeffekten bei quantitativen Merkmalen. Unpubl. Diplomarbeit: Universität Dortmund. Finney, D. J. (1945). The fractional replication of factorial experiments. Ann. Eugenics, 12, 291–301 (Correction 15, 276). Finney, D. J. (1956). Cross-over designs in bioassay. Proc. Roy. Soc., B, 145, 42–60. Fisher, R. A. (1940). An examination of the different possible solutions of a problem in incomplete blocks. Ann. Eugenics, 10, 52–75. Fisher, R. A. (1942). The theory of confounding in factorial experiments in relation to the theory of groups. Ann. Eugenics, 11, 341–353. Fisher, R. A. (1945). A system of confounding for factors with more than two alternatives, giving completely orthogonal cubes and higher powers. Ann. Eugenics, 12, 283–290. Fisher, R. A. (1949). The Design of Experiments, 5th ed. London: Hafner. Fisher, R. A. and F. Yates (1957). Statistical Tables for Biological, Agricultural and Medical Research, 5th ed. New York: Hafner. Fletcher, D. J., S. M. Lewis and J. N. S. Matthews (1990). Factorial designs for crossover clinical trials. Statist. Medicine, 9, 1121–1129. Folks, J. L. and O. Kempthorne (1960). The efficiency of blocking in incomplete block designs. Biometrika, 47, 273–283. Fries, A. and W. G. Hunter (1980). Minimum aberration 2k−p designs. Technometrics, 22, 601–608. Fujii, Y. (1976). An upper bound of resolution in symmetrical fractional factorial designs. Ann. Statist., 4, 662–667. Ghosh, S. (1981). On some new search designs for 2m factorial experiments. J. Statist. Plann. Inf., 5, 381–389. Ghosh, S. (1982). Robustness of BIBD against the unavailability of data. J. Statist. Plann. Inf., 6, 29–32. Ghosh, S. and C. Burns (2001). Two general classes of search designs for factor screening experiments with factors at three levels. Metrika, 54, 1–7. Ghosh, S. and C. Burns (2002). Comparison of four new classes of search designs. Austr. New Zealand J. Statist., 44, 357–366. Ghosh, S. and L. Teschmacher (2002). Comparisons of search designs using search probabilities. J. Statist. Plann. Inf., 104, 439–458. Gill, P. S. (1992). Balanced change-over designs for autocorrelated observations. Austral. J. Statist., 34, 415–420. Good, I. J. (1958). The interaction algorithm and practical Fourier analysis. J. Roy. Statist. Soc., B., 20, 361–372 (Addendum 22, 372–375). Graybill, F. A. (1969). Matrices with Applications in Statistics. Belmont, CA: Wadsworth.

REFERENCES

755

Graybill, F. A. and R. B. Deal (1959). Combining unbiased estimators. Biometrics, 15, 543–550. Graybill, F. A. and D. L. Weeks (1959). Combining inter-block and intra-block information in balanced incomplete block designs. Ann. Math. Statist., 30, 799–805. Grundy, P. M. and M. J. R. Healy (1950). Restricted randomization and quasi-Latin squares. J. Roy. Statist. Soc., B., 12, 286–291. Gupta, B. C. (1990). A survey of search designs for 2m factorial experiments. In: Probability, Statistics and Design of Experiment, R. R. Bahadur (ed.). New Delhi: Wiley Eastern. pp. 329–345. Gupta, S. and R. Mukerjee (1989). A Calculus for Factorial Arrangements. Berlin: Springer. Harshbarger, B. (1947). Rectangular lattices. Va. Agri. Expt. Stat. Memoir, 1, 1–26. Harshbarger, B. (1949). Triple rectangular lattices. Biometrics, 5, 1–13. Harshbarger, B. (1951). Near balance rectangular lattices. Va. J. Sci., 2, 13–27. Harshbarger, B. and L. L. Davis (1952). Latinized rectangular lattices. Biometrics, 8, 73–84. Hartley, H. O. and J. N. K. Rao (1967). Maximum-likelihood estimation for the mixed analysis of variance model. Biometrika, 54, 93–108. Hartley, H. O. and C. A. B. Smith (1948). The construction of Youden squares. J. Roy. Statist. Soc., B, 10, 262–263. Hedayat, A. and K. Afsarinejad (1975). Repeated measurement designs I. In: A Survey of Statistical Design and Linear Models. J. N. Srivastava, (ed.). Amsterdam: North Holland, pp. 229–242. Hedayat, A. and K. Afsarinejad (1978). Repeated measurement designs II. Ann. Statist., 6, 619–628. Hedayat, A. and P. W. M. John (1974). Resistant and susceptible BIB designs. Ann. Statist., 2, 148–158. Hedayat, A. S. and D. Majumdar(1984). A-Optimal incomplete block designs for controltest treatment comparisons. Technometrics, 26, 363–370. Hedayat, A. and E. Seiden (1970). F -square and orthogonal F -square design: A generalization of Latin square and orthogonal Latin square design. Ann. Math. Statist., 41, 2035–2044. Hedayat A. S. and W. D. Wallis (1978). Hadamard matrices and their applications. Ann. Statist., 6, 1184–1238. Hedayat A. S. and M. Yang (2003). Universal optimality of balanced uniform crossover designs. Ann. Statist., 31, 978–983. Hedayat, A. S., M. Jacroux and D. Majumdar (1988). Optimal designs for comparing test treatments with controls (with discussion). Statist. Sci., 4, 462–491. Hedayat, A., D. Raghavarao and E. Seiden (1975). Further contributions to the theory of F -squares design. Ann. Statist., 3, 712–716. Hedayat, A. S., N. J. A. Sloane and J. Stufken (1999). Orthogonal Arrays: Theory and Applications. New York: Springer. Hemmerle, W. J. and H. O. Hartley (1973). Computing maximum likelihood estimates for the mixed A. O. V. model using the W -transformation. Technometrics, 15, 819–831. Hinkelmann, K. (1964). Extended group divisible partially balanced incomplete block designs. Ann. Math. Statist., 35, 681–695.

756

REFERENCES

Hinkelmann, K. (1997). The Kempthorne-parameterization for asymmetrical factorials. Technical Report 97-7, Department of Statistics, Virginia Tech, Blacksburg, VA. Hinkelmann, K. (2001). Remembering Oscar Kempthorne (1919–2000). Statist. Sci., 16, 169–183. Hinkelmann, K. and O. Kempthorne (1963). Two classes of group divisible partial diallel crosses. Biometrika, 50, 281–291. Hinkelmann, K. and O. Kempthorne (1994). Design and Analysis of Experiments. Volume 1. Introduction to Experimental Design. New York: Wiley. Hoblyn, T. N., S. C. Pearce and G. H. Freeman (1954). Some considerations in the design of successive experiments in fruit plantations. Biometrics, 10, 503–515. Holcomb, D. R., D. C. Montgomery and W. M. Carlyle (2003). Analysis of supersaturated designs. J. Qual. Tech., 35, 13–27. Hotelling, H. (1944). Some improvements in weighing and other experimental techniques. Ann. Math. Statist., 15, 297–306. Huang, W.-C. L. (1989). Applications of the Chinese remainder theorem to the construction and analysis of confounding systems and randomized fractional replicates for mixed factorial experiments. Unpublished Ph.D. dissertation, Virginia Polytechnic Institute and State University, Blacksburg, VA. Jacroux, M. (1984). Upper bounds for efficiency factors of block designs. Sankhya, B, 46, 263–274. Jacroux, M. (1987). On the determination and construction of MV-optimal block designs for comparing test treatments with a standard treatment. J. Statist. Plann. Inf., 15, 205–225. Jarrett, R. G. (1983). Definitions and properties for m-concurrence designs. J. Roy. Statist. Soc., B., 64, 1–10. Jarrett, R. G. and W. B. Hall (1978). Generalized cyclic incomplete block designs. Biometrika, 65, 397–401. John, J. A. (1966). Cyclic incomplete block designs. J. Roy. Statist. Soc., B., 28, 345–360. John, J. A. (1969). A relationship between cyclic and PBIB designs. Sankhya, B, 31, 535–540. John, J. A. (1973). Factorial experiments in cyclic designs. Ann. Statist, 6, 188–194. John, J. A. (1987). Cyclic Designs. London: Chapman and Hall. John, J. A. and A. M. Dean (1975). Single replicate factorial experiments in generalized cyclic designs: I. Symmetrical arrangements. J. Roy. Statist. Soc., B, 37, 63–71. John, J. A. and J. A. Eccleston (1986). Row-column α-designs. Biometrika, 73, 301–306. John, J. A. and M. H. Quenouille (1977). Experiments: Design and Analysis, 2nd ed. New York: Macmillan. John, J. A. and T. M. F. Smith (1972). Two factor experiments in nonorthogonal designs. J. Roy. Statist. Soc., B, 34, 401–409. John, J. A. and D. J. Street (1992). Bounds for the efficiency factor of row-column designs. Biometrika, 79, 658–661. John, J. A. and D. J. Street (1993). Bounds for the efficiency factor of row-column designs (amendment). Biometrika, 80, 712–713. John, J. A. and E. R. Williams (1995). Cyclic and Computer Generated Designs, 2nd ed. London: Chapman and Hall.

REFERENCES

757

John, J. A., F. W. Wolock and H. A. David (1972). Cyclic designs. Appl. Math. Series 62. National Bureau Standards: Washington, D.C. John, P. W. M. (1966). An extension of the triangular association scheme to three associate classes. J. Roy. Statist. Soc., B, 28, 361–365. John, P. W. M. (1976). Robustness of balanced incomplete block designs. Ann. Statist., 4, 956–959. Jones, B. and M. G. Kenward (2003). Design and Analysis of Cross-over Trials, 2nd ed. Boca Raton, FL: Chapman and Hall/CRC. Kackar, R. N. (1985). Off-line quality control, parameter design, and the Taguchi method. J. Qual. Tech., 17, 176–188. Kageyama, S. (1972). On the reduction of associate classes for certain PBIB designs. Ann. Math. Statist., 43, 1528–1540. Kempthorne, O. (1947). A simple approach to confounding and fractional replication in factorial experiments. Biometrika, 34, 255–272. Kempthorne, O. (1948). The factorial approach to the weighing problem. Ann. Math. Statist., 19, 238–245. Kempthorne, O. (1952). Design and Analysis of Experiments. New York: Wiley. Kempthorne, O. (1953). A class of experimental designs using blocks of two plots. Ann. Math. Statist., 24, 76–84. Kempthorne, O. (1956). The efficiency factor of an incomplete block design. Ann. Math. Statist., 27, 846–849. Kempthorne, O. (1984). Statistical methods and science. In: W. G. Cochran’s Impact on Statistics, P.S.R.S. Rao and J. Sedransk (eds.). New York: Wiley, pp. 287–308. Kempthorne, O. and W. T. Federer (1948a). The general theory of prime-power lattice designs. I. Introduction and designs for p n varieties in blocks of p plots. Biometrics, 4, 54–79. Kempthorne, O. and W. T. Federer (1948b). The general theory of prime-power lattice designs. II. Designs for p n varieties in blocks of p s plots and in squares. Biometrics, 4, 109–121. Kershner, R. P. and W. T. Federer (1981). Two-treatment crossover designs for estimating a variety of effects. J. Am. Statist. Assoc., 76, 612–618. Khare, M. and W. T. Federer (1981). A simple construction procedure for resolvable incomplete block designs for any number of treatments. Biom. J., 23, 121–132. Kiefer, J. (1958). On the nonrandomized optimality and randomized nonoptimality of symmetrical designs. Ann. Math. Statist., 29, 675–699. Kiefer, J. (1975a). Construction and optimality of generalized Youden designs. In: A Survey of Statistical Design and Linear Models, J. N. Srivastava (ed.). Amsterdam: North-Holland, pp. 333–353. Kiefer, J. (1975b). Balanced block designs and generalized Youden designs. I. Construction (Patchwork). Ann. Statist., 3, 109–118. Kishen, K. and J. N. Srivastava (1959). Mathematical theory of confounding in asymmetrical and symmetrical factorial designs. J. Indian Soc. Agric. Statist., 11, 73–110. Kramer, C. Y. and R. A. Bradley (1957). Intra-block analysis for factorials in two-associate class group divisible designs. Ann. Math. Statist., 28, 349–361 (Addenda, 29, 933–935). Kshirsagar, A. M. (1966). Balanced factorial designs. J. Roy. Statist. Soc., B, 28, 559–567.

758

REFERENCES

Kunert, J. (1984). Optimality of balanced uniform repeated measurements designs. Ann. Statist., 12, 1006–1017. Kunert, J. (1985). Optimal repeated measurements designs for correlated observations and analysis by weighted least squares. Biometrika, 72, 375–389. Kunert, J. (1997). On the use of the factor-sparsity assumption to get an estimate of the variance in saturated designs. Technometrics, 39, 81–90. Kunert, J. and J. Stufken (2002). Optimal crossover designs in a model with self and mixed carryover effects. J. Am. Statist. Assoc., 97, 898–906. Kurkjian, B. and M. Zelen (1962). A calculus for factorial arrangements. Ann. Math. Statist., 33, 600–619. Kurkjian, B. and M. Zelen (1963). Applications of the calculus of factorial arrangements. I. Block and direct product designs. Biometrika, 50, 63–73. Kusumoto, K. (1965). Hyper cubic designs. Wakayama Med. Rept., 9, 123–132. Lenth, R. V. (1989). Quick and easy analysis of unreplicated factorials. Technometrics, 31, 469–473. Lentner, M. and T. Bishop (1993). Experimental Design and Analysis, 2nd ed. Blacksburg, VA: Valley Book. Leon, R. V., A. C. Shoemaker and R. N. Kackar (1987). Performance measure independent of adjustment: An explanation and extension of Taguchi’s signal to noise ratio. Technometrics, 29, 253–265. Lewis, S. M. (1982). Generators for asymmetrical factorial experiments. J. Statist. Plann. Inf., 6, 59–64. Li, J. C. R. (1944). Design and statistical analysis of some confounded factorial experiments. Iowa Agri. Expt. Stat. Res. Bull. 333, 452–492. Li, R. and D. K. J. Lin (2002). Data analysis in supersaturated designs. Statist. Prob. Lett., 59, 135–144. Li, R. and D. K. J. Lin (2003). Analysis methods for supersaturated design: Some comparisons. J. Data Sci., 1, 249–260. Lin, D. K. J. (1993). A new class of supersaturated designs. Technometrics, 35, 28–31. Lin, D. K. J. and N. R. Draper (1992). Projection properties of Plackett and Burman designs. Technometrics, 34, 423–428. Lin, D. K. J. and N. R. Draper (1993). Generating alias relationships for two-level Plackett and Burman designs. Comp. Statist. Data Anal., 15, 147–157. Lin, P. K. (1986). Using the Chinese remainder theorem in constructing confounded designs in mixed factorial experiments. Comm. Statist.—Theory Methods, 15, 1389– 1398. Liu, M. and R. Zhang (2000). Construction of E(s 2 ) optimal supersaturated designs using cyclic BIBDs. J. Statist. Plann. Inf., 91, 139–150. Lorenzen, T. J. and M. A. Wincek (1992). Blocking is simply fractionation. GM Research Publication 7709. Lucas, H. L. (1957). Extra-period Latin-square change-over designs. J. Dairy Sci., 40, 225–239. Majumdar, D. (1996). Optimal and efficient treatment-control designs. In: Handbook of Statistics 13, S. Ghosh and C. R. Rao (eds.). Amsterdam: Elsevier Science, pp. 1007– 1053.

REFERENCES

759

Majumdar, D. and W. Notz (1983). Optimal incomplete block designs for comparing treatments with a control. Ann. Statist., 11, 258–266. Margolin, B. H. (1968). Orthogonal main-effect 2n 3m designs and two-factor aliasing. Technometrics, 10, 559–573. Margolin, B. H. (1969). Resolution IV fractional factorial designs. J. Roy. Statist. Soc., B, 31, 514–523. Mason, J. M. and K. Hinkelmann (1971). Change-over designs for testing different treatment factors at several levels. Biometrics, 27, 430–435. Masuyama, M. (1965). Cyclic generation of triangular PBIB designs. Rep. Stat. Appl. Res. Un. Japan, Sci. Engrs., 12, 73–81. Mathew, T., B. K. Sinha and L. Zhou (1993). Some statistical procedures for combining independent tests. J. Am. Statist. Assoc., 88, 912–919. Mathon, R. and A. Rosa (1996). 2-(v, k, λ) designs of small order. In: The CRC Handbook of Combinatorial Designs, C. J. Colbourn and J. H. Dinitz (eds.). Boca Raton, FL: CRC, pp. 3–41. Matthews, J. N. S. (1987). Optimal crossover designs for the comparison of two treatments in the presence of carryover effects and autocorrelated errors. Biometrika, 74, 311–320. Montgomery, D. C. (1991). Introduction to Statistical Quality Control, 2nd ed. New York: Wiley. Montgomery, D. C. (1999). Experimental design for product and process design and development. Statistician, 48, 159–173. Morgan, J. P. (1996). Nested designs. In: Handbook of Statistics, 13, S. Ghosh and C. R. Rao (eds.). Amsterdam: Elsevier Science, pp. 939–976. Morgan, J. P. and R. A. Bailey (2000). Optimal design with many blocking factors. Ann. Statist., 28, 553–577. Muller, E.-R. (1966). Balanced confounding in factorial experiments. Biometrika, 53, 507–524. Myers, R. H., A. I. Khuri and G. Vining (1992). Response surface alternatives to the Taguchi robust parameter design approach. Am. Statist., 46, 131–139. Nair, K. R. (1944). The recovery of inter-block information in incomplete block designs. Sankhya, 6, 383–390. Nair, K. R. (1951). Rectangular lattices and partially balanced incomplete block designs. Biometrics, 7, 145–154. Nair, V. N. (1986). Testing in industrial experiments with ordered categorical data. Technometrics, 28, 283–291. Neyman, J., K. Iwaszkiewicz and S. Kolodziejczyk (1935). Statistical problems in agricultural experimentation. Suppl. J. Roy. Statist. Soc., 2, 107–154. Nguyen, N.-K. (1996). An algorithmic approach to constructing supersaturated designs. Technometrics, 38, 69–73. Nguyen, N.-K. and E. R. Williams (1993). An algorithm for constructing optimal resolvable row-column designs. Austral. J. Statist., 35, 363–370. Ogasawara, M. (1965). A necessary condition for the existence of regular and symmetric PBIB designs. Inst. Stat. Mimeo Series 418: Chapel Hill, NC. Ogawa, J. (1974). Statistical Theory of the Analysis of Experimental Designs. New York: Dekker.

760

REFERENCES

Ohnishi, T. and T. Shirakura (1985). Search designs for 2n factorial experiments. J. Statist. Plann. Inf., 11, 241–245. Paley R. E. A. C. (1933). On orthogonal matrices. J. Math. Phys., 12, 311–320. Paterson, L. J. (1983). Circuits and efficiency in incomplete block designs. Biometrika, 70, 215–225. Patterson, H. D. (1951). Change-over trials. J. Roy. Statist. Soc., B, 13, 256–271. Patterson, H. D. (1952). The construction of balanced designs for experiments involving sequences of treatments. Biometrika, 39, 32–48. Patterson, H. D. and H. L. Lucas (1959). Extra-period change-over designs. Biometrics, 15, 116–132. Patterson, H. D. and H. L. Lucas (1962). Change-over designs. North Carolina Agr. Exp. Station Tech. Bull., 147. Patterson, H. D. and R. Thompson (1971). Recovery of inter-block information when block sizes are unequal. Biometrika, 58, 545–554. Patterson, H. D. and E. R. Williams (1976). A new class of resolvable incomplete block designs. Biometrika, 63, 83–92. Pearce, S. C. (1953). Field experimentation with fruit trees and other perennial plants. Commonwealth Bur. Horticulture Plantation Crops, Tech. Comm., 23. Pearce, S. C. (1960). Supplemented balance. Biometrika, 47, 263–271 (Corrigenda 48, 475). Pearce, S. C. (1963). The use and classification of non-orthogonal designs (with discussion). J. Roy. Statist. Soc., A, 126, 353–377. Pearce, S. C. (1964). Experimenting with blocks of natural size. Biometrics, 20, 699–706. Phadke, M. S. (1982). Quality engineering using design of experiments. ASA Proceedings, Section on Statistical Education, 11–20. Alexandria, VA: American Statistical Association. Plackett, R. L. (1946). Some generalizations in the multifactorial design. Biometrika, 33, 328–332. Plackett, R. L. and J. P. Burman (1946). The design of optimum multifactorial experiments. Biometrika, 33, 305–325. Prvan, T. and D. J. Street (2002). An annotated bibliography of application papers using certain classes of fractional factorial and related designs. J. Statist. Plann. Inf., 106, 245–269. Pukelsheim, F. (1993). Optimal Design of Experiments. New York: Wiley. Quenouille, M. H. (1953). The Design and Analysis of Experiments. London: Griffin. Raghavarao, D. (1960). A generalization of group divisible designs. Ann. Math. Statist., 31, 756–771. Raghavarao, D. (1971). Construction and Combinatorial Problems in Design of Experiments. New York: Wiley. Raghavarao, D. (1986). Random balance designs. In: Encyclopedia of Statistical Science. S. Kotz, N. L. Johnson and C. B. Read (eds.). New York: Wiley, pp. 505–506. Raghavarao, D. (1990). Cross-over designs in industry. In: Statistical Design and Analysis of Industrial Experiments. S. Ghosh (ed.). New York: Dekker, pp. 517–530. Raghavarao, D. and K. Chandrasekhararao (1964). Cubic designs. Ann. Math. Statist., 35, 389–397.

REFERENCES

761

Raghavarao, D. and S. K. Thartare (1967). An inequality for doubly balanced incomplete block designs. Calcutta Statist. Assoc. Bull., 16, 37–39. Raghavarao, D. and S. K. Thartare (1970). A new series of doubly balanced designs. Calcutta Statist. Assoc. Bull., 19, 95–96. Raktoe, B. L. (1969). Combining elements from distinct finite fields in mixed factorials. Ann. Math. Statist., 40, 498–504. Raktoe, B. L. (1970). Generalized combining of elements from finite fields. Ann. Math. Statist., 41, 1763–1767. Raktoe, B. L., A. Hedayat and W. T. Federer (1981). Factorial Designs. New York: Wiley. Raktoe, B. L., A. A. Rayner and D. O. Chalton (1978). On construction of confounded mixed factorial and lattice designs. Austral. J. Statist., 20, 209–218. Rao, C. R. (1946a). Hypercubes of strength d leading to confounded designs in factorial experiments. Bull. Calcutta Math. Soc., 38, 67–78. Rao, C. R. (1946b). Difference sets and combinatorial arrangements derivable from finite geometries. Proc. Nat. Inst. Sci. India, 12, 123–135. Rao, C. R. (1947a). General methods of analysis for incomplete block designs. J. Am. Statist. Assoc., 42, 541–561. Rao, C. R. (1947b). Factorial experiments derivable from combinatorial arrangements of arrays. J. Roy. Statist. Soc., B, 9, 128–139. Rao, C. R. (1950). The theory of fractional replication in factorial experiments. Sankhya, 10, 81–86. Rao, V. R. (1958). A note on balanced designs. Ann. Math. Statist., 29, 290–294. Rashed, D. H. (1984). Designs for multiple comparisons of control versus test treatments. Unpublished Ph.D. dissertation, Virginia Polytechnic Institute and State University, Blacksburg, VA. Ratkowsky, D. A., M. A. Evans and J. R. Alldredge (1993). Cross-over Experiments— Design, Analysis and Application. New York: Dekker. Ray-Chaudhuri, D. K. (1965). Some configurations in finite projective spaces and partially balanced incomplete block designs. Can. J. Math., 17, 114–123. Riedwyl, H. (1998). Modifying and using Yates’ algorithm. Statist. Papers, 39, 41–60. Roy, J. and K. R. Shah (1962). Recovery of interblock information. Sankhya, A, 24, 269–280. Roy, P. M. (1953–1954). Hierarchical group divisible incomplete block designs with m-associate classes. Sci. Culture, 19, 210–211. Russell, K. G., J. A. Eccleston and G. J. Knudsen (1981). Algorithms for the construction of (M, S)-optimal block designs and row-column designs. J. Statist. Comp. Simul., 12, 93–105. r Proprietary Software Release 8.2. Cary, NC: SAS SAS Institute (1999–2001). SAS Institute. Satterthwaite, F. E. (1946). An approximate distribution of estimates of variance components. Biometrics, 2, 110–114. Satterthwaite, F. E. (1959). Random balance experimentation. Technometrics, 1, 111–137. Seely, J. and D. Birkes (1984). Parametrizations and resolution IV. In: Experimental Design, Statistical Models, and Genetic Statistics. K. Hinkelmann (ed.). New York: Dekker, pp. 77–94.

762

REFERENCES

Sen, M. and R. Mukerjee (1987). Optimal repeated measurements designs under interaction. J. Statist. Plann. Inf., 17, 81–91. Senn, S. J. (1993). Cross-over Trials in Clinical Research. Chichester, U.K.: Wiley. Seshradi, V. (1963). Combining unbiased estimators. Biometrics, 19, 163–170. Shah, B. V. (1958). On balancing in factorial experiments. Ann. Math. Statist., 29, 766– 779. Shah, B. V. (1959). A generalization of partially balanced incomplete block designs. Ann. Math. Statist., 30, 1041–1050. Shah, B. V. (1960). On a 5 × 22 factorial experiment. Biometrics, 16, 115–118. Shah, K. R. (1964). Use of inter-block information to obtain uniformly better estimators. Ann. Math. Statist., 35, 1064–1078. Shah, K. R. (1970). On the loss of information in combined inter- and intra-block estimation. J. Am. Statist. Assoc., 65, 1562–1564. Shah, K. R. (1977). Analysis of designs with two-way elimination of heterogeneity. J. Statist. Plann. Inf., 1, 207–216. Sharma, V. K. (1977). Change-over designs with complete balance for first and second residual effects. Canadian J. Statist., 5, 121–132. Shing, C.-C. and K. Hinkelmann (1992). Repeated measurement designs for factorial treatments when high order interactions are negligible. J. Statist. Plann. Inf., 31, 81–91. Shirakura, T. (1991). Main effect plus one or two plans for 2m factorials. J. Statist. Plann. Inf., 27, 65–74. Shirakura, T. and T. Ohnishi (1985). Search designs for 2m factorials derived from balanced arrays of strength 2( + 1) and AD-optimal search designs. J. Statist. Plann. Inf., 11, 247–258. Shirakura, T., T. Takahashi and J. N. Srivastava (1996). Searching probabilities for nonzero effects in search designs for the noisy case. Ann. Statist., 24, 2560–2568. Shrikhande, S. S. (1950). The impossibility of certain symmetrical balanced incomplete block designs. Ann. Math. Statist., 21, 106–111. Shrikhande, S. S. (1952). On the dual of some balanced incomplete block designs. Biometrics, 8, 66–72. Shrikhande, S. S. (1960). Relations between certain incomplete block designs. In: Contributions to Probability and Statistics. I. Olkin (ed.). Stanford, CA: Stanford University Press, pp. 388–395. Shrikhande, S. S. (1965). On a class of partially balanced incomplete block designs. Ann. Math. Statist., 36, 1807–1814. Shrikhande, S. S. and D. Raghavarao (1963). A method of constructing incomplete block designs. Sankhya, A, 25, 399–402. Sihota, S. S. and K. S. Banerjee (1981). On the algebraic structures in the construction of confounding plans in mixed factorial designs on the lines of White and Hultquist. J. Am. Statist. Assoc., 76, 996–1001. Sitter, R. R., J. Chen and M. Feder (1997). Fractional resolution and minimum aberration in blocking factorial designs. Technometrics, 39, 382–390. Srivastava, J. N. (1965). Optimal balanced 2m fractional factorial designs. In: S. N. Roy Memorial Volume. Chapel Hill, NC: University of North Carolina and Calcutta: Indian Statistical Institute.

REFERENCES

763

Srivastava, J. N. (1975). Designs for searching non-negligible effects. In: A Survey of Statistical Design and Linear Models. J. N. Srivastava (ed.). Amsterdam: North-Holland, pp. 507–519. Srivastava, J. N. (1984). Sensitivity and revealing power: Two fundamental statistical criteria other than optimality arising in discrete experimentation. In: Experimental Design, Statistical Models, and Genetic Statistics. K. Hinkelmann (ed.). New York: Dekker, pp. 95–117. Srivastava, J. N. (1993). Nonadditivity in row-column designs. J. Comb. Inform. System Sci., 18, 85–96. Srivastava, J. N. (1996). A critique of some aspects of experimental design. In: Handbook of Statistics, 13. S. Ghosh and C. R. Rao (eds.). Amsterdam: Elsevier Science, pp. 309–341. Srivastava, J. N. and D. A. Anderson (1970). Some basic properties of multidimensional partially balanced designs. Ann. Math. Statist., 41, 1438–1445. Srivastava, J. N. and D. V. Chopra (1971). Balanced optimal 2m fractional factorial designs of resolution V, m ≤ 6. Technometrics, 13, 257–269. Srivastava, J. N. and D. V. Chopra (1973). Balanced arrays and orthogonal arrays. In: A Survey of Combinatorial Theory. J. N. Srivastava, F. Harary, C. R. Rao, G.-C. Rota, and S. S. Shrikhande (eds.). Amsterdam: North-Holland, pp. 411–428. Srivastava, J. N. and S. Ghosh (1976). A series of balanced 2m factorial designs of resolution V which allow search and estimation of one extra unknown effect. Sankhya, B, 38, 280–289. Srivastava, J. N. and S. Ghosh (1977). Balanced 2m factorial designs of resolution V which allow search and estimation of one extra unknown effect, 4 ≤ m ≤ 8. Comm. Statist.—Theor. Meth., 6, 141–166. Srivastava, J. N. and S. Ghosh (1996). On nonorthogonality and nonoptimality of Addelman’s main-effect plans satisfying the condition of proportional frequencies. Statist. Prob. Lett., 26, 51–60. Srivastava, J. N. and B. C. Gupta (1979). Main effect plan for 2m factorials which allow search and estimation of one unknown effect. J. Statist. Plann. Inf., 3, 259–265. Srivastava, J. N. and Y. C. Wang (1998). Row-column designs: How and why nonadditivity may make them unfit for use. J. Statist. Plann. Inf., 73, 277–315. Srivastava, R., V. K. Gupta and A. Dey (1990). Robustness of some designs against missing observations. Comm. Statist., A, 19, 121–126. Street, A. P. and D. J. Street (1987). Combinatorics of Experimental Design. Oxford: Clarendon. Stufken, J. (1987). A-optimal block designs for comparing test treatments with a control. Ann. Statist., 15, 1629–1638. Stufken, J. (1991). On group divisible treatment designs for comparing test treatments with a standard treatment in blocks of size 3. J. Statist. Plann. Inf., 28, 205–211. Stufken, J. (1996). Optimal crossover designs. In: Handbook of Statistics, Vol. 13. S. Ghosh and C. R. Rao (eds.). Amsterdam: Elsevier Science, pp. 63–90. Sun, D. X., C. F. J. Wu and Y. Y. Chen (1997). Optimal blocking schemes for 2n and 2n−p designs. Technometrics, 39, 298–307. Surendran, P. U. (1968). Association matrices and the Kronecker product of designs. Ann. Math. Statist., 39, 676–680.

764

REFERENCES

Taguchi, G. (1986). Introduction to Quality Engineering: Designing Quality into Products and Processes. Tokyo: Asian Productivity Organization. Taguchi, G. (1987). System of Experimental Design: Engineering Methods to Optimize Quality and Minimize Costs. White Plains, NY: UNIPUB/Kraus Internation. Taguchi, G. (1993). Taguchi Methods—Design of Experiments. Dearborn, MI: American Supplies Institute. Takeuchi, K. (1962). A table of difference sets generating balanced incomplete block designs. Rev. Intern. Statist. Inst., 30, 361–366. Tang, B. and C. F. J. Wu (1997). A method for constructing supersaturated designs and its E s 2 optimality. Canadian J. Statist., 25, 191–201. Tharthare, S. K. (1963). Right angular designs. Ann. Math. Statist., 34, 1057–1067. Tharthare, S. K. (1965). Generalized right angular designs. Ann. Math. Statist., 36, 1051– 1062. Thompson, W. A. (1956). A note on the balanced incomplete block designs. Ann. Math. Statist., 27, 842–846. Thompson, W. A., Jr. (1958). A note on PBIB design matrices. Ann. Math. Statist., 29, 919–922. Thompson, W. A. (1962). The problem of negative estimates of variance components. Ann. Math. Statist., 33, 273–289. Tjur, T. (1990). A new upper bound for the efficiency factor of a block design. Austral. J. Statist., 32, 231–237. van der Waerden, B. L. (1966). Algebra. Berlin: Springer. Vartak, M. N. (1955). On an application of Kronecker product of matrices to statistical designs. Ann. Math. Statist., 26, 420–438. Vartak, M. N. (1959). The non-existence of certain PBIB designs. Ann. Math. Statist., 30, 1051–1062. Venables, W. N. and J. A. Eccleston (1993). Randomized search strategies for finding optimal or near-optimal block and row-column designs. Austral J. Statist., 35, 371–382. Vining, G. G. and R. H. Myers (1990). Combining Taguchi and response surface philosophies: A dual response approach. J. Quality Tech., 22, 38–45. Voss, D. T. (1986). On generalizations of the classical method of confounding to asymmetrical factorial experiments. Comm. Statist., A, 15, 1299–1314. Wald, A. (1943). On the efficient design of statistical investigations. Ann. Math. Statist., 14, 134–140. Webb, S. R. (1964). Characterization of non-orthogonal incomplete factorial designs. In: Appendix B of ARL 65-116, Part I. Aerospace Research Laboratories: US. Air Force, Wright-Patterson Air Force Base, Dayton, OH. Webb, S. R. (1968). Saturated sequential factorial designs. Technometrics, 10, 535–550. Webb, S. R. (1971). Small incomplete factorial experiment designs for two- and three-level factors. Technometrics, 13, 243–256 (Correction 15, 951). Westfall, P. H., S. S. Young and D. K. J. Lin (1998). Forward selection error control in the analysis of supersaturated designs. Statist. Sinica, 8, 101–117. Whitaker, D., E. R. Williams and J. A. John (1997). CycDesigN: A Package for the Computer Generation of Experimental Design. Canberra: CSIRO.

REFERENCES

765

White, D. and R. A. Hultquist (1965). Construction of confounding plans for mixed factorial designs. Ann. Math. Statist., 36, 1256–1271. Williams, E. J. (1949). Experimental designs balanced for the estimation of residual effects of treatments. Austral. J. Sci. Res., A, 2, 149–168. Williams, E. J. (1950). Experimental designs balanced for pairs of residual effects. Austral. J. Sci. Res., A, 3, 351–363. Williams, E. R. and J. A. John (2000). Updating the average efficiency factor in α-designs. Biometrika, 87, 695–699. Williams, E. R. and M. Talbot (1993). ALPHA+: Experimental Designs for Variety Trials. Design User Manual. Canberra: CSIRO and Edinburgh: SASS. Wolfinger, R. D. and R. D. Tobias (1998). Joint estimation of location, dispersion, and random effects in robust design. Technometrics, 40, 62–71. Wolock, F. W. (1964). Cyclic designs. Unpublished Ph.D. dissertation, Virginia Polytechnic Institute, Blacksburg, VA. Worthley, R. and K. S. Banerjee (1974). A general approach to confounding plans in mixed factorial experiments when the number of a factor is any positive integer. Ann. Statist., 2, 579–585. Wu, C. F. J. (1993). Construction of supersaturated designs through partially aliased interactions. Biometrika, 80, 661–669. Wu, C. F. J. and Y. Chen (1992). A graph-aided method for planning two-level experiments when certain interactions are important. Technometrics, 34, 162–175. Wu, C. F. J. and M. Hamada (2000). Experiments—Planning, Analysis, and Parameter Design Optimization. New York: Wiley. Wu, H. and C. F. J. Wu (2002). Clear two-factor interactions and minimum aberration. Ann. Statist., 30, 1496–1511. Yamada, S., Y. T. Ikebe, H. Hashiguchi and N. Niki (1999). Construction of three-level supersaturated design. J. Statist. Plann. Inf., 81, 183–193. Yamada, S. and D. K. J. Lin (1999). Three-level supersaturated designs. Statist. Probab. Lett., 45, 31–39. Yamada, S. and T. Matsui (2002). Optimality of mixed-level supersaturated design. J. Statist. Plann. Inf., 104, 459–468. Yamamoto, S., Y. Fuji and N. Hamada (1965). Composition of some series of association algebras. J. Sci. Hiroshima Univ., A, 29, 181–215. Yates, F. (1935). Complex experiments. J. Roy. Statist. Soc., B, 2, 181–223. Yates, F. (1936a). Incomplete randomized blocks. Ann. Eugenics, 7, 121–140. Yates, F. (1936b). A new method of arranging variety trial involving a large number of varieties. J. Agri. Sci, 26, 424–455. Yates, F. (1937a). A further note on the arrangement of variety trials: Quasi-Latin squares. Ann. Eugenics, 7, 319–332. Yates, F. (1937b). The design and analysis of factorial experiments. Imp. Bur. Soil Sci. Tech. Comm., 35, 1–95. Yates, F. (1939). The recovery of inter-block information in variety trials arranged in three-dimensional lattices. Ann. Eugenics, 9, 136–156. Yates, F. (1940a). The recovery of inter-block information in balanced incomplete block designs. Ann. Eugenics, 10, 317–325.

766

REFERENCES

Yates, F. (1940b). Lattice squares. J. Agric. Sci., 30, 672–687. Youden, W. J. (1937). Use of incomplete block replications in estimating tobacco-mosiac virus. Contr. Boyce Thompson Inst., 9, 41–48. Youden, W. J. (1940). Experimental designs to increase accuracy of greenhouse studies. Contr. Boyce Thompson Inst., 11, 219–228. Youden, W. J., O. Kempthorne, J. W. Tukey, G. E. P. Box and J. S. Hunter (1959). Discussion of “Random balance experimentation” by F. E. Satterthwaite and “The applications of random balance designs” by T. A. Budne. Technometrics, 1, 157–184. Zahn, D. A. (1975). Modifications of revised critical values for the half-normal plot. Technometrics, 17, 189–200. Zelen, M. (1958). The use of group divisible designs for confounded asymmetrical factorial arrangements. Ann. Math. Statist., 29, 22–40. Zhang, R. and D. K. Park (2000). Optimal blocking of two-level fractional factorial designs. J. Statist. Plann. Inf., 91, 107–121. Zhou, L. and T. Mathew (1993). Combining independent tests in linear models. J. Am. Statist. Assoc., 80, 650–655. Zoellner, J. A. and O. Kempthorne (1954). Incomplete block designs with blocks of two plots. Iowa State College Agri. Expt. Stat. Res. Bull., 418.

Author Index

Abel, R. J. R., 656 Abraham, B., 605 Addelman, S., 517, 567, 583, 585–586, 593 Adhikary, B., 154 Afsarinejad, K., 685, 687, 689, 692, 698, 710–711, 713 Aggarwal, K. R., 178, 185, 188, 505, 727 Alldredge, J. R., 695, 706 Anderson, D. A., 14 Atkinson, A. C., 50 Autrey, K. M., 684 Bailey, R. A., 238 Baksalary, J. K., 103 Balaam, L. N., 694 Banerjee, K. S., 491, 576 Baumert, L. D., 576 Bechhofer, R. E., 199, 201, 203–205, 208, 212 Berenblut, I. I., 693, 713 Beyer, W. H., 104 Bhattacharya, C. G., 38 Bhattacharya, K. N., 158 Birkes, D., 543 Bisgaard, S., 554 Bishop, T., 81 Blaisdell, E. A., 691 Booth, K. H. V., 598–599 Bose, R. C., 3, 14, 92–93, 104, 108, 119–120, 123–124, 127–130, 137–138, 156, 158, 167, 173, 394, 409, 449, 582 Box, G. E. P., 516, 543, 597, 605, 638, 640 Bradley, J. V., 688 Bradley, R. A., 503 Brouwer, A. E., 656

Bulutoglu, D. A., 600 Burman, J. P., 575–576 Burns, C., 615 Bush, K. A., 582 Calinski, T., 22 Calvin, L. D., 102 Cannon, C. Y., 684 Carlyle, W. M., 607 Carmichael, R. D., 716 Chakravarti, I. M., 547–549, 726–727 Chalton, D. O., 491 Chandrasekhararao, K., 146 Chang, C.-T., 180, 182, 185, 189, 505 Chang, L.-C., 166–167 Chatterjee, K., 615 Chen, H., 522, 524, 551, 553–555 Chen, J., 554 Chen, Y., 518 Chen, Y. Y., 554 Cheng, C.-S., 49, 522, 524, 551, 553–555, 600–601, 687, 693, 695, 699 Chipman, H., 605 Chopra, D. V., 549, 726–727 Clatworthy, W. H., 130, 132, 137, 160, 162, 164–167, 210–211, 503 Cochran, W. G., 91, 104, 231, 684–685 Cohen, A., 79, 81 Colbourn, C. J., 656 Connor, W. S., 130, 138, 156 Constantine, G., 198 Corbeil, R. R., 40–41 Cotter, S. C., 454–455 Cotton, J. W., 685


767

768 Cox, D. R., xxii, 196, 598–599 Cox, G. M., 91, 104, 231 Daniel, C., 558–559, 563, 595 Das, M. N., 196 David, H. A., 153, 169 Davies, O. L., 559 Davis, A. W., 691 Davis, L. L., 90, 240 Deal, R. B., 78 Dean, A. M., 457, 462, 464, 477–478, 480 Dénes, J., 688 Deng, L.-Y., 615 Dey, A., 103, 531, 576, 585, 593 Dinitz, J. H., 656 Donev, A. N., 50 Draper, N. R., 577 Dunnett, C. W., 199, 212 Eccleston, J. A., 230, 232–233, 237–238 Ehrenfeld, S., 48 Engel, J., 640–641 Evans, M. A., 695, 706 Fan, J., 606 Fang, K.-T., 603 Feder, M., 554 Federer, W. T., 114–115, 194–195, 531, 537, 539, 654, 671, 695 Fichtner, A., 195 Finney, D. J., 509, 685 Fisher, R. A., 3, 73, 104, 118, 254, 289, 374, 394, 403, 565 Fletcher, D. J., 715 Folks, J. L., 16 Freeman, G. H., 196 Fries, A., 518, 536 Fuji, Y., 154 Fujii, Y., 534 Ghosh, S., 103, 585, 610, 612, 614–617, 630–632 Gill, P. S., 708 Golomb, S. W., 576 Good, I. J., 264 Graybill, F. A., 76, 78 Grundy, P. M., 338 Gupta, B. C., 610, 612 Gupta, S., 480, 494–495, 499 Gupta, V. K., 103 Hall, M., Jr., 576 Hall, W. B., 193, 237, 691

AUTHOR INDEX

Hamada, M., 518, 522, 529, 545, 634, 637, 648 Hamada, N., 154 Harshbarger, B., 3, 90, 240, 679, 682 Hartley, H. O., 40, 91 Hashiguchi, H., 603 Healy, M. J. R., 338 Hedayat, A., 102, 197, 199, 201–202, 211, 231, 531, 537, 539, 576, 685, 698–699, 710–711, 713, 719, 724 Hemmerle, W. J., 40 Hinkelmann, K., xxii, xix, 147, 176, 180, 182, 185, 481, 493, 505, 713–714 Hoblyn, T. N., 196 Holcomb, D. R., 607 Hotelling, H., 576 Huang, W.-C. L., 491 Hultquist, R. A., 473, 481 Hunter, J. S., 516, 597 Hunter, W. G., 518, 536 Ikebe, Y. T., 603 Iwaszkiewicz, K., 51 Jacroux, M., 47, 197, 199, 201, 205, 211 Jarrett, R. G., 47, 193, 237 John. J. A., 47, 90, 153, 169, 190, 192, 213, 232–233, 237–240, 448, 454–455, 457–458, 462, 464, 477–478, 480, 683, 709 John, P. W. M., 102, 144 Jones, B., 695 Kackar, R. N., 633–635, 640 Kageyama, S., 22, 177, 187 Keedwell, A. D., 688 Kempthorne, O., xxii, xix, 2, 16, 26, 46, 147, 153, 170, 176, 256, 337, 394, 492–493, 551, 567, 576–577, 583, 585–586, 593, 597, 654, 657, 665, 669, 671 Kenward, M. G., 695 Kershner, R. P., 695 Khare, M., 114–115, 194–195 Khuri, A. I., 637, 641, 648 Kiefer, J., 48, 50, 231, 712 Kishen, K., 394, 506 Knudsen, G. J., 238 Kolodziejczyk, S., 51 Kramer, C. Y., 503 Kshirsagar, A. M., 493–494, 499 Kunert, J., 563, 698–699, 710–713 Kurkjian, B., 172, 448–449, 468, 493–495 Kusumoto, K., 149

769

AUTHOR INDEX

Lenth, R. V., 562–563 Lentner, M., 81 Leon, R. V., 640 Lewis, S. M., 479, 715 Li, J. C. R., 494, 505–506, 728–731, 733 Li, R., 605–607 Lin, D. K. J., 576–577, 597, 599–600, 603–607, 615 Lin, P. K., 491 Liu, C.-W., 166 Liu, M., 601–603 Liu, W.-R., 166–167 Lorenzen, T. J., 551 Lucas, H. L., 685, 690–691 Ma, C.-X., 603 Majumdar, D., 197–199, 201–202, 211 Margolin, B. H., 543, 576, 593 Mason, J. M., 713 Masuyama, M., 166 Mathew, T., 81 Mathon, R., 115, 119 Matsui, T., 598 Matthews, J. N. S., 708, 715 McGilchrist, C. A., 233 Mesner, D. M., 120, 127–130, 173 Meyer, R. D., 605, 640 Montgomery, D. C., 607, 633–634 Morgan, J. P., 238 Mukerjee, R., 480, 494–495, 499, 615, 710 Muller, E.-R., 506 Myers, R. H., 637, 641, 648 Nair, K. R., 3–4, 119, 124, 137, 655, 680–682 Nair, V. N., 639 Neyman, J., 51 Nguyen, N.-K., 238–239, 601 Niki, N., 603 Notz, W., 211 Ogasawara, M., 154 Ogawa, J., 22 Ohnishi, T., 612–614, 616 Paley, R. E. A. C., 576 Park, D. K., 555 Paterson, L. J., 47 Patterson, H. D., 40–41, 190–191, 682, 685, 690–691 Pearce, S. C., 13, 72, 196 Phadke, M. S., 639 Plackett, R. L., 567, 575–576 Prvan, T., 509

Pukelsheim, F., 48, 50 Puri, P. D., 103 Quenouille, M. H., 709 Raghavarao, D., 92–93, 102, 104, 110, 113, 115, 119, 143, 146, 154, 165, 231, 597, 685, 691 Raktoe, B. L., 481, 491, 531, 537, 539 Rao, C. R., 4, 104, 394, 406, 547, 582 Rao, J. N. K., 40 Rao, V. R., 72 Rashed, D. H., 205–208 Ratkowsky, D. A., 695, 706 Ray-Chaudhuri, D. K., 166 Rayner, A. A., 491 Reid, N., xxii Riedwyl, H., 264 Rosa, A., 115, 119 Roy, J., 38 Roy, P. M., 143, 154 Russell, K. G., 230, 232, 238 Sackrowitz, H. B., 79, 81 SAS Institute, 13, 39, 52, 267 Satterthwaite, F. E., 387, 596–597 Searle, S. R., 40–41 Seely, J., 543 Seiden, E., 231 Sen, M., 710 Senn, S. J., 685 Seshradi, V., 78 Shah, B. V., 9, 149, 493–494, 735 Shah, K. R., 38, 78–79, 155, 232 Sharma, V. K., 709 Shimamoto, T., 123, 137 Shing, C.-C., 714 Shirakura, T., 612–614, 616, 630–631 Shoemaker, A. C., 640 Shrikhande, S. S., 74, 92–93, 137, 158–159, 166–167 Sihota, S. S., 491 Singh, T. P., 727 Sinha, B. K., 81 Sitter, R. R., 554 Sloane, N. J. A., 576, 719, 724 Smith, C. A. B., 91 Smith, T. M. F., 454–455 Srivastava, J. N., 14, 213, 506, 549, 585, 608–610, 612, 614, 616–617, 625, 630–631, 726–727 Srivastava, R., 103 Steinberg, D. M., 522, 555 Street, A. P., 130

770 Street, D. J., 130, 509, 683 Stufken, J., 197, 205, 211, 576, 686, 695, 699, 710–713, 719, 724 Sun, D. X., 522, 554–555 Surendran, P. U., 172 Taguchi, G., 633–634, 636, 638 Takahashi, T., 630–631 Takeuchi, K., 115 Talbot, M., 192 Tamhane, A. C., 199, 201, 203–205, 208, 212 Tang, B., 601 Teschmacher, L., 630–632 Thartare, S. K., 102, 151, 153 Thompson, R., 40–41 Thompson, W. A., 40, 56, 120 Tjur, T., 47 Tobias, R. D., 648 Tukey, J. W., 597 van der Waerden, B. L., 481 Vartak, M. N., 142, 147, 172, 176–177 Venables, W. N., 238 Vijayan, K., 605 Vining, G., 637, 641, 648 Voss, D. T., 485 Wald, A., 48 Wallis, W. D., 576 Wang, Y. C., 213 Webb, S. R, 517, 595

AUTHOR INDEX

Weeks, D. L., 78 Westfall, P. H., 605 Whitaker, D., 192 White, D., 473, 481 Williams, E. J., 688, 708–709 Williams, E. R., 47, 90, 190–192, 213, 238–240, 682–683 Wilson, K. B., 543 Wincek, M. A., 551 Wolfinger, R. D., 648 Wolock, F. W., 153, 169 Worthley, R., 491 Wu, C.-F., 687, 693, 695, 699 Wu, C. F. J., 518, 522, 529, 545, 554, 600–601, 634, 637, 648 Wu, H., 522 Yamada, S., 597–599, 603–604 Yamamoto, S., 154 Yang, M., 699 Yates, F., 3–4, 23, 71, 104, 114, 118, 167, 238, 263, 338, 394, 492, 505, 576, 649, 665–666, 671, 729, 732 Youden, W. J., 3, 91, 597 Young, S. S., 605 Zahn, D. A., 559, 563 Zelen, M., 172, 448–449, 468, 493–495 Zhang, R., 555, 601–603 Zhou, L., 81 Zoellner, J. A., 2, 153

Subject Index

Absorb option, 58 Additivity in the broad sense, 4, 22 in the strict sense, 16 Adjusted treatment total, 7 Algebra(s), 129 linear associative, 127–129 Algorithm interchange, 239 inverse, 264 Yates, 263 Aliasing, see Structure, alias Alias, 510 Analysis of balanced incomplete block design, 74 combined, 4, 13, 26–27, 61–63, 66, 76, 79, 82, 131, 134, 156, 311, 323, 387 incomplete block design, 3 interblock, 3–4, 23, 26, 87, 303, 421 intrablock, 3–4, 13–14, 46, 54, 58, 63–64, 74, 79, 82, 131, 135, 311, 323, 387, 413 randomization, 16–18, 22, 39 Analysis of variance (ANOVA), 12–13, 18, 90, 96–97, 215, 265, 278, 281, 303, 310, 320, 367, 640–641 B|T-, 12, 15, 21, 316–317, 332–333, 349, 382–383, 386, 427–428, 652, 661–663 T|B-, 12, 15, 21, 314–316, 326, 331–333, 381–382, 386, 421, 442, 651–652, 661–662 disconnected incomplete block design, 15 sequential, 216

table(s), see also B|T-, T|Bfor disconnected incomplete block design, 15 for factorial experiment, 420 for incomplete block design, 12, 52, 61, 298 for row-column design, 216 Array(s) balanced (B), 178–180, 549, 614, 724, 726–727 combinatorial, 547 common, 636 cross, 636–642 generating, 190 inner, 636, 639 Kronecker product, 547, 636 noise, 639 orthogonal (OA), 158, 164, 405–406, 547–549, 593, 636, 709, 724, 727 of strength 2, 164, 582 outer, 636, 639 partially balanced, 726 reduced, 190 single, 636–637 square, 655 Association, 123 matrix, 120, 121, 127, 133–135, 173–174 Association scheme(s), 120, 123, 127, 133, 137, 154, 158–159, 169–172, 174–176, 208, 231 cubic, 146, 670 cyclic, 141, 153


771

772 Association scheme(s) (Continued) extended group-divisible (EGD), 147, 176 generalized group-divisible, 143–144 right-angular, 153 triangular, 144 group-divisible, 138, 160–164, 208 hypercubic, 149, 177 Latin square, 670 L2 , 140, 166 Li , 140 rectangular, 142, 176 right-angular, 151 triangular, 139, 165 Autocorrelation, 707 Autoregression, first-order, 707

Balanced incomplete block (BIB) design(s), 2–4, 9, 52, 71–73, 89–92, 95, 100–120, 136, 159, 168, 172, 175–180, 187, 196–197, 201, 207, 211, 281, 313, 602, 649, 653–654, 690 affine resolvable, 93–95 α-resolvable, 92–93, 96 complement of, 113 dual of, 158, 166 irreducible, 113 reinforced, 196 residual, 114 resistant, 98–102 globally, 99–102 locally, 99–100 resolvable, 93, 96–97, 314 robust, 103 susceptible, 98–99 symmetrical, 74, 105, 114, 711 Balanced treatment incomplete block (BTIB) design(s), 199–204, 208 optimal, 211–212 R-type, 202 S -type, 201–203 Balancedness, 687, 695, 706–707 Bar chart, 561 Block(s) comparisons, 23, 283, 322 effect(s), 3–4, 13, 16, 22, 279–281, 300, 399 eliminating treatments, 13 ignoring treatments, 13 initial, 105–106, 110, 167–169, 193, 458, 463–464, 602 size(s), 1–2, 14, 279 unequal, 13, 99, 189, 192

SUBJECT INDEX

Bound lower, 601 upper, 47–48, 75, 192, 233, 682–683 Branching column, 599 Chinese remainder theorem, 491 Collapsing of factor levels, 567, 583, 586–587, 591 Component(s) pure block, 301 treatment, 301, 421 Comparison(s), see also Contrast(s) block, 23, 283, 322 pure, 421 orthogonal, 248 post-hoc, 54, 64, 67 treatment, 3–4, 23, 43, 171, 236, 284, 376, 430 Confidence coefficient, 212 interval(s), 212 Confounded, completely, 417–419 Confounding, 279, 288, 368, 650 complete, 312, 380, 417 double, 335–336, 379 in squares, 336 partial, 312–314, 318, 324, 327, 331–334, 343, 351, 375, 380–382, 492, 653 system(s) of, 323, 329 efficiency of, 323 relationship(s), 511–513 system(s) of, 258, 264, 283, 287–289, 292–293, 298–300, 327, 362, 366, 370, 374–375, 378–380, 398–405, 430, 433, 443–444, 452–453, 459–461, 466, 473–474, 480, 491, 549–550, 564, 655–659, 713 constructing, 457, 491 detecting, 291 symmetrical, 467 Connectedness, 14, 103, 232 Contrasts, see also Comparison(s) orthogonal, 243 Control factor, 634–636, 639–641 quality, 633 off-line, 633 treatment, see Treatment, control Criterion, see also Optimality average-variance, 49 chi-squared (χ 2 ), 598 maximum, 604 determinant, 49

SUBJECT INDEX

E(s 2 ), 601 smallest eigenvalue, 49 Crossover design(s), 684–690, 715 2-treatment, 693–695 balanced, 687–688, 699, 702, 707–708 completely, 709 for residual effects, 687 minimal, 688–689 minimal completely, 709 minimal strongly, 691–692 partially, 691 strongly, 687, 691–692, 699, 706, 710 strongly uniform, 693 totally, 712–713 uniform, 692, 698–699 construction of, 688, 708 extra-period, 691, 699, 702, 712 orthogonal, 709 uniform, 693 Cyclic development, 107–109, 193–194 Dealiasing, 545, 569, 577, 609 Defining equation(s), see Defining relationship Defining relationship, 511–513, 545. See also Identity relationship Design(s), see also Experiment(s); Plan(s) A-optimal, 199 alpha (α), 190–192 2-occurrence class, 190 3-occurrence class, 191 analysis of balanced incomplete block, 74 complete block, 666–669 disconnected incomplete block, 15 incomplete block, 3 binary, 71, 99, 194 block balanced, 712 extended, 196, 231 central composite, 637–639 changeover, see Crossover design(s) column, 213, 230–231, 237 complete block, randomized, see Design(s), randomized complete block completely randomized (CRD), 43, 46, 263, 278, 366 component, 213, 230–234, 237 computer search, 598 confounded, 376, 407 partially, 314, 317, 325 connected, 43, 98, 230 crossover, see Crossover design(s) cyclic, 169–170 construction of, 167

773 fractional, 169 disconnected, 281, 460 dual, 159 efficient, 89, 239, 708 equireplicate, 31, 34, 71, 99 error control, 16, 43, 230, 241, 262, 272, 366, 452, 685 experimental, 1–3, 241 construction of, 164 factorial, 633. See also Factorial(s) asymmetrical, 547 balanced (BFD), 491–506 blocked fractional, 549, 555 fractional, 507–509, 513, 517, 531, 547–549, 636, 640 minimum aberration blocked fractional, 555 resolution of a fractional, 518 symmetrical, 394 two-level, 241 foldover, 543–544 partial, 545 generalized cyclic (GC/n), see also Design(s), incomplete block, generalized cyclic randomized block, 231 Youden, 231, 712 generator, 201–205, 212 highly fractionated, 531 incomplete block, 1–4, 9, 13, 23, 38, 43, 46, 50, 91, 119, 123, 131, 169, 172, 189, 194, 281, 310, 318, 323, 413, 448, 491, 649, 657, 669 ANOVA table of, 12 balanced, see Balanced incomplete block (BIB) design(s) balanced treatment, see Balanced treatment incomplete block (BTIB) design(s) binary, 155 disconnected, 14–15, 462 disconnected resolved, 417 doubly balanced, 102 equireplicate, 120, 155, 452 generalized cyclic, 193, 237, 448 Latinized, 90 partially balanced, see Partially balanced incomplete block (PBIB) design(s) partially balanced treatment, see Partially balanced treatment incomplete block (PTBIB) design(s) proper, 120, 155, 452 resolvable, 189, 194–195, 649 resolved, 418 symmetrical, 73

774 Design(s) (Continued) incomplete Latin square, 91, 231 Kronecker product, 172–178, 502, 529, 548–549, 583, 636 Latin square, 3, 50–51, 90, 263, 688. See also Square(s), Latin Latin square type, 1, 190, 706 lattice(s), 3–4, 114, 119, 649–669. See also Lattice(s) maximum resolution, 518–520, 534–535 minimum aberration, 520–522 near-optimal, 192 nonorthogonal, 563 optimal, 49, 52, 211, 695–699, 711–713 orthogonal, 271, 563 paired comparison, 169 process, 633 proper, 31, 34, 71, 99 quasi-factorial, 3, 114, 119, 649 randomized complete block (RCBD), 1, 4, 43, 46, 51, 238, 263, 267, 278, 281, 300, 366–367, 666, 669 random variables, 16, 18 repeated measurements, see Crossover design(s) replicate, 238 residual, 172 resolution III, 516, 527–528, 537–539, 543–546, 582. See also Resolution, III resolution IV, 516, 522, 537–539, 543. See also Resolution, IV resolution V, 516, 537, 546. See also Resolution, V resolvable, 114, 189, 238, 653 resolved, 424 response surface, 49, 637 robust parameter, 634 row, 213, 230–231, 234, 237 row-column, 190, 213–216, 230–234, 237, 240, 684–685, 712 doubly incomplete, 230 Latinized, 240 nested, 238 regular, 230–231, 237 resolvable, 238, 682 sampling, 272, 685 saturated, 539, 546 search, 608–614, 631–632 constructing, 614 sequential construction of, 614 sequential, 544 square, see Square(s) supersaturated, 596–608 2-level, 598–599, 603

SUBJECT INDEX

3-level, 603 balanced incomplete block (BIBD)-based, 601 Hadamard type, 599–601 mixed level, 598 supplemented balance, 196–197, 200 switch-back, 693 symmetric, 50 treatment, 241, 272, 452 treatment-control, 205 treatment-control comparison, 198 unconfounded, 323, 376 variance-balanced, 98–99, 103 weighing, 576 Differences mixed, 108 pure, 108 symmetrically repeated, 107–109 Difference set(s), 105–108, 112 Effect(s) block, 3–4, 13, 16, 22, 281, 300, 399 fixed, 4 random, 4 carryover, 685–686, 712 mixed, 710 self, 710, 713 clear, 518, 539, 543 strongly, 518 column, 232 direct, 685, 697–698, 708–713 dispersion, 640 estimable, 281 fixed, 4, 706–707 main, 241–242, 247, 250–251, 254, 260–262, 265, 281, 287, 324, 359–360, 365, 526, 546, 713 period, 686, 706 random, 707 residual, 685–686, 697–699, 706–710, 713 first-order, 708–709 second-order, 708–709 row, 232 simple, 242, 249 sparsity, 558, 605 subject, 686, 707 treatment, 3, 16, 22, 197, 232, 399, 685–686, 706 Effective number of replications, 79 Efficiency(ies), 3, 168, 197, 334, 699 factor(s), 43–48, 75, 103, 129–131, 134, 171–172, 190–194, 198, 232, 236–238, 498, 501, 682 relative, 43

775

SUBJECT INDEX

Eigenvalue(s), 8, 129–131, 155, 170, 198, 233 Eigenvector(s), 8, 44, 233, 424 Equation(s) Aitken, 27–28, 32 normal (NE), 5, 10, 28, 134, 213–216, 696 combined, 36 interblock, 23–27 reduced (RNE), 5–10, 13, 27, 74, 119–121, 131, 170, 214–216, 238, 320–321, 413–414, 452, 494, 539, 697–699 prediction, 641 Error control, 213 experimental, 1, 5, 14, 189, 214, 279–281, 323, 610 homogeneity of, 302 interblock, 665–666 intrablock, 3, 76, 665 observational, 5, 214, 281, 610 pseudostandard, 562 rate(s), familywise, 605 reduction, 213 simultaneous margin of, 562 variance, 76, 379 whole-plot, 639 Estimate(s), see Estimation; Estimator(s) Estimation capacity, 522 combined, 31, 36 index, 522 interblock, 31 intrablock, 31 maximum-likelihood, 39, 648 restricted (residual) maximum-likelihood (REML), 40–41, 648 Estimator(s) Aitken, 38 best linear unbiased (BLUE), 416–418, 421–424 intrablock, 3 combined, 34, 38–39, 77–78, 155, 333, 427–429, 651–653, 672 interblock, 34–36, 303, 651–653 intrablock, 34–39, 78, 133, 155, 303, 651–653, 672 residual (restricted) maximum-likelihood (REML), 40–42 Shah, 155 Yates, 37–39, 77–78, 96, 136 Experiment(s), see also Design(s); Factorial(s) 23 , 324 24 , 327 25 , 335

2n , 329 32 , 377 33 , 363, 367–370, 377 35 , 402 3n , 363, 366 confounded, 303 cross-array, 640–641 exploratory-type, 278 factorial, 114, 279, 323, 657 analysis of, 262 confounded, 657 field plot, 1 robust-design, 633–637, 642, 646 rotation, 685 search, 617–619 analysis of, 617 single-array, 641 split-plot type, 595 supersaturated, 604–606 analysis of, 604 Experimental unit, 2, 74 Experimentation, sequential, 648 Factor(s) adjustment, 639–641, 648 blocking, 189, 551–553 control, 634–636, 639–641 design, 634–636, 639–641, 646–648 environmental, 634 expanding, 587 highest common, 458 noise, 634–636, 640–641, 646–648 pseudo, 431, 435, 441–443, 463, 467, 491, 660 qualitative, 362 quantitative, 362 random, 646 scarcity of active, 596, 605 treatment, 551 Factor level(s), collapsing of, 567, 583, 586–587, 591 Factorial(s), see also Design(s); Experiment(s) 22 , 650 23 , 267, 272, 279, 282, 293–295, 298, 304–306, 309–311, 325–326, 337–338, 343, 511, 714 24 , 293, 296, 338, 350–351 26 , 441 (22 )3 , 440 2n , 264, 278, 283, 288–289, 293, 303, 312, 329, 513 32 , 359, 388 33 , 366, 371 34 , 405

776 Factorial(s) (Continued) 3n , 368, 374, 380, 524 42 , 431 43 , 434–436, 441–442 62 , 443–446, 460 6n , 443–444 pn , 394, 398–401, 405–406, 410–412, 428 (pm )n , 431 pmn , 431 s n , 431, 447, 452–454, 457, 465, 529, 558, 564–565 analysis of, 265, 343, 366, 412 asymmetrical, 465–467, 474, 477, 480, 483–485, 488, 505, 529, 558, 567, 582, 586, 595 balanced, 728–734 confounding in, 466 calculus, 448 fractional, 264, 362, 507–509, 525, 547–549, 576, 715 asymmetrical, 548 blocking in, 549 nonorthogonal, 546 regular, 517 resolution of, 516 mixed, see Factorial(s), asymmetrical symmetrical, 409, 431, 467, 477, 480, 488, 517, 529, 564, 583, 593–595, 653 confounding for, 447 unreplicated, 558 2n , 558 Field(s) finite, 481, 716 Galois (GF), 112, 480–483, 716–718, 724 theory, 112, 431 Foldover principle, 543–545, 577 Fraction(s) foldover, 544 regular, 531 Fractional replicate(s), 513, 517, 524, 531 replication, 509–510, 516, 529 Frequencies nonproportional, 594 proportional, 567, 583–586 Function(s) confounded, 417 estimable, 8, 11, 38, 52, 75, 197, 215, 230, 424, 541, 550–551, 699 linearly independent, 462 of treatment effects, 10, 35 linearly independent, 421 loss, 635, 639 minimum, 129

SUBJECT INDEX

penalty, 606 smoothly clipped absolute deviation (SCAD), 606 Gauss–Markov theorem, 3, 10 Generalized inverse(s), 7–10, 30, 61, 131, 198, 541 Generator(s), 401 of the intrablock subgroup, 407 Geometry(ies) Euclidean, 162, 409, 721–723 finite, 158, 162, 409, 721 projective, 162, 409, 721–724 Group(s) Abelian, 105, 110, 128, 289, 458 additive, 108–112 cyclic, 688 Guardrails, 559–560 Half-normal plot(s), 558–560 Harmonic mean, 45–46 Heterogeneity eliminating, in two directions, 3, 89, 213 two-way elimination of, 2, 162 Hypercube of strength 2, 406 of strength d, 406 Ideal(s), 481–484 Identity relationship(s), 513–518, 525–526, 529–531, 544, 550, 554, 558, 637, 641 Information, 280, 323, 334 combined intra- and interblock, 154, 333, 349, 386, 426, 653, 658 function, 48 interblock, 3, 26–27, 46, 77, 303, 308, 314, 332, 385, 422–426, 707 recovery of, 3–4, 22–23, 40 intersubject, 707 intrablock, 3, 26, 48, 308, 314, 330, 350, 375, 381, 421–422, 426, 653, 658, 666 intrasubject, 707 matrix, see Matrix, information relative, 337, 501 total relative loss in, 493 within-row-and-column, 336 Interaction(s), 241–243, 247, 251, 254, 260–262, 265, 281, 286, 297, 359, 360–365, 449 1-factor, 297 2-factor, 250, 280, 292, 324, 337, 361, 373–374, 379, 395, 403–405, 434, 492, 536, 539, 546, 608–609

777

SUBJECT INDEX

3-factor, 250, 280, 337, 368, 378–379, 434, 567–568, 608 n-factor, 395, 493 component(s), 366–369, 376, 412, 422, 493, 658, 713 modified, 487 confounded, 279, 282, 401, 407, 425, 462, 489 design-noise factor, 647 direct-residual treatment, 710 generalized (GI), 258–260, 285–286, 364, 396–399, 408, 433, 514, 525, 551 high-order, 280, 508, 516, 608–609 independent, 285–287, 292, 399–402, 407, 433–434, 468, 514, 529, 551 independent confounded, 374 low-order, 287, 403, 513, 608 mixed, 530–531 negligible, 279, 614 nonnegligible, 608 partial, 261–265 Intrablock subgroup (IBSG), 289–291, 370, 374, 379, 400–401, 405–408, 434–435, 484, 550, 565 Lagrangian multipliers, 419 Lattice(s) analysis of, 657–659 balanced, 653–656, 659 efficiency of, 669 higher-dimensional, 654, 657 one-restrictional, 654–655, 658, 668–669, 682 two-dimensional (square), 654–655, 670, 679, 682–683 three-dimensional (cubic), 654–656, 659, 662 quadruple, 654–656 quintuple, 656 rectangle(s), 678–679 rectangular, 3, 194, 654–655, 679 near-balanced, 654 one-restrictional, 654 simple, 654, 680 triple, 654, 681–682 septuple, 656 sextuple, 656 simple, 655–656, 670 triple, 655–656, 659, 662 two-restrictional, see Lattice square Lattice square, 239, 654, 671, 675–678 balanced, 654, 671–672 semibalanced, 654, 671 unbalanced, 654, 675

Least squares, 28 generalized, 27 means (LSM), 58, 69, 230, 271, 706 direct, 707 residual, 707 ordinary, 4 penalized, 605–606 theory of, 3 Likelihood function, 40–42 log-, 41 Loss, 633–635 average, 636 expected, 635 function, 635, 639 total relative, 501 Matrix(ces) association, 120–121, 127, 173–174 C , see Matrix(ces), information circulant, 170, 459 coefficient, 25, 121 completely symmetric, 50 concordance, 72, 155, 173, 452 concurrence, 72, 190, 196, 199 cyclic, 460 design, 238, 612 design-model, 413 Hadamard, 575–576, 593, 599, 600, 603 cyclic, 600 noncyclic, 600 normalized, 576 seminormalized, 576, 599–600 incidence, 5, 71, 113–114, 158, 173, 177, 187, 452 observation-block, 473 period-residual effect, 686 period-treatment, 686 treatment-block, 473 information, 7, 48, 129, 197–198, 232, 236–238, 695–698, 708 transformation, 244, 251 variance-covariance, 11, 24, 75, 197, 383, 424, 428, 496, 539, 549 Maximum likelihood (ML), 64–66, 648 residual (REML), 66–67, 81, 392, 648 Method direct, 180, 188 of differences, 158–160 successive diagonalizing, 115, 194–195 Taguchi, 633–637 Minimum aberration, 555 Missing plots, 13 Model(s) general linear mixed, 40

778 Model(s) (Continued) linear, 4, 262 mixed, 648 search, 608–609 location-scale, 639 mixed, 333 random effects, 40 regression, 605, 647 response surface, 647 three-part, 313 Modeling dispersion, 638, 641 dual-response, 641 location, 638, 641 Noiseless, 610, 618–619 Noisy, 625, 630 Nonorthogonality, 3, 599, 605 measure of, 599 Optimal, see also Optimality A-, 197, 211 blocking, 551 E(s 2 )-, 601–603 universally, 698, 712–713 Optimality, 48–50, 601, 693, 708, 711 A-, 49 criterion, 48–49, 197 D-, 49 E-, 49 Kiefer, 50 MS-, 238 MV-, 199 universal, 50, 712 Orthogonal series, 114 Orthogonality, 284, 583, 593, 596–597, 691, 713 adjusted, 232, 237 between-interaction, 494 deviation from, 597–598 measure of, 597–598 near-, 596–597 within-interaction, 494 Parameter(s) of the first kind, 124 of the second kind, 122–128 Parameterization of treatment responses, 245, 256, 297, 365, 410, 485 non-full rank, 366 Partially balanced incomplete block (PBIB) design(s), 3–4, 9, 52, 119, 122–127, 131,

SUBJECT INDEX

137, 158, 172–173, 207–209, 281, 467, 649, 653, 669, 691 2-associate class (PBIB/2), 129, 132, 136–137, 155–156, 168, 172, 190 cyclic, 137, 141, 157, 169 group-divisible (GD), 137–138, 156–164, 208, 502–504 L2 , 140, 166, 670 L3 , 670 Li , 140, 670 Latin square, 166 Latin square type, 137, 140, 157 regular group-divisible, 156 semiregular group-divisible, 156 singular group-divisible, 156 triangular, 137–139, 156, 165 3-associate class (PBIB/3), 169, 176, 190, 231, 234, 670 cubic, 137, 146, 670 generalized group-divisible (GGD), 137, 143 generalized triangular, 137, 144 rectangular, 137, 142, 147–148 4-associate class (PBIB/4), 680 generalized right-angular, 153 right-angular, 137, 151 resolved, 680 m-associate class (PBIB/m), 127, 131, 134 cubic, 149 cyclic, 137, 167–170, 193, 460, 691 extended group-divisible (EGD), 137, 147–148, 178, 187, 466, 493, 499–505 EGD/(2ν −1), 147, 177–187 EGD/3, 176–179, 186, 502 EGD/7, 180 generalized cyclic (GC/n), 458–460, 465, 477–478 hypercubic, 137, 149, 187–188 Partially balanced treatment incomplete block (PTBIB) design(s), 205, 208–209, 230 construction of, 208–209 type I, 205–206 type II, 207, 210–211 Partition(s), 260, 363–364, 395–396, 401, 404, 409 cyclic, 454 hierarchical, 457, 461 mixed, 488, 536 orthogonal, 255, 393 pure, 487, 536 Performance measure, 640 Plan(s) Fisher, 403, 407, 564, 576

779

SUBJECT INDEX

main effect(s) (MEP), 564, 608–609 auxiliary, 586–589, 593 constructing, 585–586 Fisher, 564–566, 577, 597 nonorthogonal, 594 Plackett–Burman, 575–578, 597–599 orthogonal (OMEP), 564–569, 575–577, 582–595, 614 saturated, 566, 594 plus one (MEP.1), 611–614 resolution V plus one, 614–617 saturated, 594, 603, 611 weak MEP.1, 612 one-at-a-time, 594 Polynomial irreducible, 718 orthogonal, 361 primitive, 115, 719 Primitive mark, 717 polynomial, 115, 719 root, 112–113, 115, 717 Procedure(s) ANOVA, 342 computational, 52 Dunnett’s, 199, 212 maximum likelihood (ML), 63 randomization, 16 residual (restricted) maximum likelihood (REML), 63, 342 SAS, see SAS Satterthwaite, 387, 392 Shah, 79 variable selection, 605 Yates, 37–38, 61–63, 70, 77, 386, 664 Product Kronecker, 172, 177, 244, 251, 450, 469 symbolic direct (SDP), 172–173, 186, 448, 468 symbolic inner (SIP), 179–180 Projection, 577 Proportionality condition, 585 Pseudofactor(s), see Factor(s), pseudo Randomization, 595 restricted, 338, 707 test(s), see Test(s), randomization theory, 212 Regression, 605, 641, 647 all-subsets, 605 linear, 605 Repeated measurements, see Crossover design(s) Replication groups, 89, 96, 107

Residue class, 193 Resolution, 516 2R, 517 2R + 1, 517 III, 516, 536 nonorthogonal, 594 orthogonal, 564, 582 III.1, 611, 614 IV, 516, 536, 577 V, 516, 536, 614 V.1, 614–617 VII, 608 maximum, 518, 520 upper bound of, 534 Resolvability, 92 Response surface, 648 Revealing power, 610 Ring(s), 128 finite, 468, 481–482 Robustness, 103 SAS PROC FACTEX, 293, 325, 338, 350, 370, 435, 445, 469, 474, 511, 517, 520, 527, 531 PROC GLM, 52, 58, 81, 216, 267, 277, 311, 338, 342, 350, 387, 391, 569, 618, 699 PROC MIXED, 39, 52, 63, 79–81, 267, 278, 311, 338, 342, 350, 387, 392, 707 PROC VARCOMP, 61 Search probability, 630–632 Selection stepwise variable, 605–606 subset, 606 variable, 605 Signal-to-noise (S/N) ratio(s), 638 Spectral decomposition, 8 Software, 192 Square(s) frequency, 231 Latin, 50–51, 231, 655, 688 mutually orthogonal (MOLS), 166, 656, 675, 682, 708–709 orthogonal, 691 replicated, 239 Williams, 688, 699–700, 708 Youden, 50, 91, 107, 221, 688, 711, 714 Statistical inference, 16, 51, 215 Structure alias, 373, 513–516, 520, 526–530, 567–569, 576 autocorrelation error, 707 factorial orthogonal (OFS), 452–454, 457–461, 473, 495

780

SUBJECT INDEX

Structure (Continued) treatment, 713–715 variance–covariance, 206, 708 Subclass numbers, unequal, 3 Subgroup, 290 intrablock (ISBG), 289–291, 370, 374, 379 Submodel(s), 617 Subsampling, 272 Supermodel, 609 Symbolic direct multiplication, 185. See also Product, symbolic direct

test, 196, 205 whole-plot, 639 Trial animal feeding, 685 pharmaceutical, 685 clinical, 685

Target value, 635–636, 639, 648 Test(s) F -, 266, 302, 311, 384, 391 randomization, 22, 266, 302, 384 significance, 79 Time series, 707 Transformation, 640 Treatment(s) combination(s), 242, 247, 252 comparison(s), 4, 8, 38, 368 component, 421 contrast(s), see Treatment(s), comparison(s) control, 196 effect(s), 3, 16, 22, 197, 232 eliminating blocks, 13 responses, 256–258 parameterization of, see Parameterization, of treatment responses

Variance average, 46, 168, 171, 659 component(s), 62, 350, 392 interblock, 25, 37 intrablock, 24, 37, 77 minimum, 38 residual, 90 within-block, see Variance, intrablock

Unbiasedness, 38 Uniform, 687 on the periods, 687 on the subjects, 687

Weights estimation of, 36, 286–387 inaccuracies of, 661, 665 interblock, 24 intrablock, 24 Word length pattern (WLP), 518, 554 Words block-defining, 553–554 treatment-defining, 551–554


Design and Analysis of Experiments

Design and Analysis of Experiments

Design and Analysis of Experiments

Design analysis of experiments